Merge branch 'master' into feat/claude-sdk-block

Implements a new AI agent block that bridges Claude Agent SDK with AutoGPT's tool ecosystem:

- Uses Claude Agent SDK for native agent capabilities
- Discovers and executes connected AutoGPT tool nodes
- Supports iterative agent mode with conversation management
- Proper credential validation for Anthropic API keys
- Configurable models (Claude 4.x and 3.x series)
- Zero-warning block registration and validation

Block features:
- Agent mode execution with tool discovery
- Integration with AutoGPT's execution manager
- Robust async error handling
- Support for max iterations and working directory configuration

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

autogpt_platform/backend/backend/blocks/smart_agent.py

@@ -0,0 +1,858 @@
+import logging
+import os
+import re
+from enum import Enum
+from typing import TYPE_CHECKING, Any, Literal, NamedTuple
+
+from claude_agent_sdk import query
+from claude_agent_sdk.types import (
+    AssistantMessage,
+    ClaudeAgentOptions,
+    TextBlock,
+    ToolUseBlock,
+)
+from pydantic import BaseModel, SecretStr
+
+# Avoid circular imports by importing only essential types
+# ExecutionParams, ToolInfo, and other classes will be imported dynamically when needed
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+    BlockType,
+)
+from backend.data.dynamic_fields import is_tool_pin
+from backend.data.execution import ExecutionContext
+from backend.data.model import (
+    APIKeyCredentials,
+    CredentialsField,
+    CredentialsMetaInput,
+    SchemaField,
+)
+from backend.integrations.providers import ProviderName
+from backend.util import json
+from backend.util.clients import get_database_manager_async_client
+
+if TYPE_CHECKING:
+    from backend.data.graph import Link, Node
+    from backend.executor.manager import ExecutionProcessor
+
+logger = logging.getLogger(__name__)
+
+
+class ToolInfo(BaseModel):
+    """Processed tool call information."""
+
+    tool_call: Any  # The original tool call object from LLM response
+    tool_name: str  # The function name
+    tool_def: dict[str, Any]  # The tool definition from tool_functions
+    input_data: dict[str, Any]  # Processed input data ready for tool execution
+    field_mapping: dict[str, str]  # Field name mapping for the tool
+
+
+class ExecutionParams(BaseModel):
+    """Tool execution parameters."""
+
+    user_id: str
+    graph_id: str
+    node_id: str
+    graph_version: int
+    graph_exec_id: str
+    node_exec_id: str
+    execution_context: "ExecutionContext"
+
+
+def _create_tool_response(call_id: str, content: str) -> dict[str, Any]:
+    """Create a tool response in the correct format."""
+    # Anthropics format: tool IDs typically start with "toolu_"
+    if call_id.startswith("toolu_"):
+        return {
+            "role": "user",
+            "type": "message",
+            "content": [
+                {"tool_use_id": call_id, "type": "tool_result", "content": content}
+            ],
+        }
+    # OpenAI format: tool IDs typically start with "call_".
+    # Or default fallback (if the tool_id doesn't match any known prefix)
+    return {"role": "tool", "tool_call_id": call_id, "content": content}
+
+
+class ModelMetadata(NamedTuple):
+    provider: str
+    context_window: int
+    max_output_tokens: int | None
+
+
+class AgentModel(str, Enum):
+    """Available models for the Smart Agent."""
+
+    # Claude 4.x models (latest)
+    CLAUDE_4_1_OPUS = "claude-opus-4-1-20250805"
+    CLAUDE_4_OPUS = "claude-opus-4-20250514"
+    CLAUDE_4_SONNET = "claude-sonnet-4-20250514"
+    CLAUDE_4_5_OPUS = "claude-opus-4-5-20251101"
+    CLAUDE_4_5_SONNET = "claude-sonnet-4-5-20250929"
+    CLAUDE_4_5_HAIKU = "claude-haiku-4-5-20251001"
+
+    # Claude 3.x models (stable)
+    CLAUDE_3_7_SONNET = "claude-3-7-sonnet-20250219"
+    CLAUDE_3_HAIKU = "claude-3-haiku-20240307"
+
+    @property
+    def metadata(self) -> ModelMetadata:
+        return AGENT_MODEL_METADATA[self]
+
+    @property
+    def provider(self) -> str:
+        return self.metadata.provider
+
+
+# Agent model metadata mapping
+AGENT_MODEL_METADATA = {
+    # Claude 4.x models
+    AgentModel.CLAUDE_4_1_OPUS: ModelMetadata(ProviderName.ANTHROPIC, 500000, 4096),
+    AgentModel.CLAUDE_4_OPUS: ModelMetadata(ProviderName.ANTHROPIC, 500000, 4096),
+    AgentModel.CLAUDE_4_SONNET: ModelMetadata(ProviderName.ANTHROPIC, 500000, 4096),
+    AgentModel.CLAUDE_4_5_OPUS: ModelMetadata(ProviderName.ANTHROPIC, 500000, 8192),
+    AgentModel.CLAUDE_4_5_SONNET: ModelMetadata(ProviderName.ANTHROPIC, 500000, 8192),
+    AgentModel.CLAUDE_4_5_HAIKU: ModelMetadata(ProviderName.ANTHROPIC, 200000, 4096),
+    # Claude 3.x models
+    AgentModel.CLAUDE_3_7_SONNET: ModelMetadata(ProviderName.ANTHROPIC, 200000, 4096),
+    AgentModel.CLAUDE_3_HAIKU: ModelMetadata(ProviderName.ANTHROPIC, 200000, 4096),
+}
+
+# Anthropic-only credentials for Claude models
+ClaudeCredentials = CredentialsMetaInput[
+    Literal[ProviderName.ANTHROPIC], Literal["api_key"]
+]
+
+
+def ClaudeCredentialsField() -> ClaudeCredentials:
+    return CredentialsField(
+        description="Anthropic API key for Claude Agent SDK access.",
+        discriminator="model",
+        discriminator_mapping={
+            model.value: model.metadata.provider for model in AgentModel
+        },
+    )
+
+
+# Test credentials for Claude models
+TEST_CLAUDE_CREDENTIALS = APIKeyCredentials(
+    id="test-claude-creds",
+    provider=ProviderName.ANTHROPIC,
+    api_key=SecretStr("mock-anthropic-api-key"),
+    title="Mock Anthropic API key",
+    expires_at=None,
+)
+TEST_CLAUDE_CREDENTIALS_INPUT = {
+    "provider": TEST_CLAUDE_CREDENTIALS.provider,
+    "id": TEST_CLAUDE_CREDENTIALS.id,
+    "type": TEST_CLAUDE_CREDENTIALS.type,
+    "title": TEST_CLAUDE_CREDENTIALS.title,
+}
+
+
+class SmartAgentBlock(Block):
+    """
+    A smart agent block that uses Claude Agent SDK for native agent capabilities
+    while executing AutoGPT tool nodes.
+
+    This block combines Claude's native agent functionality with AutoGPT's tool ecosystem:
+    - Uses Claude Agent SDK for core agent intelligence
+    - Discovers connected AutoGPT tool nodes like SmartDecisionMaker
+    - When Claude calls tools, executes the actual AutoGPT tool nodes
+    - Provides Claude with the tool execution results
+    """
+
+    class Input(BlockSchemaInput):
+        task: str = SchemaField(
+            description="The task for the agent to complete. Be specific about your requirements.",
+            placeholder="Analyze the data file and create a summary report with key insights...",
+        )
+        model: AgentModel = SchemaField(
+            title="Model",
+            default=AgentModel.CLAUDE_4_5_SONNET,
+            description="The model to use for the agent.",
+            advanced=False,
+        )
+        credentials: ClaudeCredentials = ClaudeCredentialsField()
+        max_iterations: int = SchemaField(
+            default=15,
+            description="Maximum number of agent iterations. Use -1 for unlimited (use carefully!).",
+            advanced=False,
+        )
+        system_prompt: str = SchemaField(
+            title="System Prompt",
+            default="You are a helpful AI assistant with access to tools. Think step by step about which tools to use to complete the task efficiently. When you have completed the objective, provide a clear summary of the results.",
+            description="System prompt to guide the agent's behavior.",
+            advanced=True,
+        )
+        working_directory: str = SchemaField(
+            default="/tmp/smart_agent",
+            description="Working directory for the agent.",
+            advanced=True,
+        )
+
+    class Output(BlockSchemaOutput):
+        result: str = SchemaField(
+            description="The final result or answer from the agent."
+        )
+        iterations_used: int = SchemaField(
+            description="Number of iterations used to complete the task."
+        )
+        tools_used: list[str] = SchemaField(
+            description="List of AutoGPT tools used during execution.",
+            default_factory=list,
+        )
+        success: bool = SchemaField(
+            description="Whether the task was completed successfully."
+        )
+        error: str = SchemaField(
+            default="", description="Error message if the task failed."
+        )
+        # Tool output pins for connecting to other blocks (like SmartDecisionMakerBlock)
+        tools: Any = SchemaField(
+            description="Tool calls output for connecting to other AutoGPT blocks."
+        )
+        conversations: list[Any] = SchemaField(
+            description="Conversation history with Claude Agent SDK.",
+            default_factory=list,
+        )
+
+    def __init__(self):
+        super().__init__(
+            id="c1a2u3d4-e5a6-g7e8-n9t0-b1l2o3c4k5d6",
+            description=(
+                "An AI agent powered by Claude Agent SDK that executes connected AutoGPT tool nodes. "
+                "Combines Claude's native agent capabilities with AutoGPT's tool ecosystem."
+            ),
+            categories={BlockCategory.AI},
+            block_type=BlockType.AI,
+            input_schema=SmartAgentBlock.Input,
+            output_schema=SmartAgentBlock.Output,
+            test_input={
+                "task": "What tools are available?",
+                "credentials": TEST_CLAUDE_CREDENTIALS_INPUT,
+                "model": AgentModel.CLAUDE_4_5_SONNET,
+            },
+            test_output=[],
+            test_credentials=TEST_CLAUDE_CREDENTIALS,
+        )
+
+    @staticmethod
+    def cleanup(s: str):
+        """Clean up block names for use as tool function names."""
+        return re.sub(r"[^a-zA-Z0-9_-]", "_", s).lower()
+
+    async def _create_tool_node_signatures(
+        self,
+        node_id: str,
+    ) -> list[dict[str, Any]]:
+        """
+        Creates function signatures for connected tools.
+        Args:
+            node_id: The node_id for which to create function signatures.
+        Returns:
+            List of function signatures for tools
+        """
+        from backend.blocks.agent import AgentExecutorBlock
+
+        db_client = get_database_manager_async_client()
+        tools = [
+            (link, node)
+            for link, node in await db_client.get_connected_output_nodes(node_id)
+            if is_tool_pin(link.source_name) and link.source_id == node_id
+        ]
+        if not tools:
+            raise ValueError("There is no next node to execute.")
+        return_tool_functions: list[dict[str, Any]] = []
+        grouped_tool_links: dict[str, tuple["Node", list["Link"]]] = {}
+        for link, node in tools:
+            if link.sink_id not in grouped_tool_links:
+                grouped_tool_links[link.sink_id] = (node, [link])
+            else:
+                grouped_tool_links[link.sink_id][1].append(link)
+        for sink_node, links in grouped_tool_links.values():
+            if not sink_node:
+                raise ValueError(f"Sink node not found: {links[0].sink_id}")
+            if sink_node.block_id == AgentExecutorBlock().id:
+                # Dynamic import to avoid circular dependency
+                from backend.blocks.smart_decision_maker import SmartDecisionMakerBlock
+
+                tool_func = (
+                    await SmartDecisionMakerBlock._create_agent_function_signature(
+                        sink_node, links
+                    )
+                )
+                return_tool_functions.append(tool_func)
+            else:
+                # Dynamic import to avoid circular dependency
+                from backend.blocks.smart_decision_maker import SmartDecisionMakerBlock
+
+                tool_func = (
+                    await SmartDecisionMakerBlock._create_block_function_signature(
+                        sink_node, links
+                    )
+                )
+                return_tool_functions.append(tool_func)
+        return return_tool_functions
+
+    async def _execute_single_tool_with_manager(
+        self,
+        tool_info: ToolInfo,
+        execution_params: ExecutionParams,
+        execution_processor: "ExecutionProcessor",
+    ) -> dict:
+        """Execute a single tool using the execution manager for proper integration."""
+        # Lazy imports to avoid circular dependencies
+        from concurrent.futures import Future
+
+        from backend.data.execution import NodeExecutionEntry
+
+        tool_call = tool_info.tool_call
+        tool_def = tool_info.tool_def
+        raw_input_data = tool_info.input_data
+
+        # Get sink node and field mapping
+        sink_node_id = tool_def["function"]["_sink_node_id"]
+
+        # Use proper database operations for tool execution
+        db_client = get_database_manager_async_client()
+
+        # Get target node
+        target_node = await db_client.get_node(sink_node_id)
+        if not target_node:
+            raise ValueError(f"Target node {sink_node_id} not found")
+
+        # Create proper node execution using upsert_execution_input
+        node_exec_result = None
+        final_input_data = None
+
+        # Add all inputs to the execution
+        if not raw_input_data:
+            raise ValueError(f"Tool call has no input data: {tool_call}")
+
+        for input_name, input_value in raw_input_data.items():
+            node_exec_result, final_input_data = await db_client.upsert_execution_input(
+                node_id=sink_node_id,
+                graph_exec_id=execution_params.graph_exec_id,
+                input_name=input_name,
+                input_data=input_value,
+            )
+
+        assert node_exec_result is not None, "node_exec_result should not be None"
+
+        # Create NodeExecutionEntry for execution manager
+        node_exec_entry = NodeExecutionEntry(
+            user_id=execution_params.user_id,
+            graph_exec_id=execution_params.graph_exec_id,
+            graph_id=execution_params.graph_id,
+            graph_version=execution_params.graph_version,
+            node_exec_id=node_exec_result.node_exec_id,
+            node_id=sink_node_id,
+            block_id=target_node.block_id,
+            inputs=final_input_data or {},
+            execution_context=execution_params.execution_context,
+        )
+
+        # Use the execution manager to execute the tool node
+        try:
+            # Get NodeExecutionProgress from the execution manager's running nodes
+            node_exec_progress = execution_processor.running_node_execution[
+                sink_node_id
+            ]
+
+            # Use the execution manager's own graph stats
+            graph_stats_pair = (
+                execution_processor.execution_stats,
+                execution_processor.execution_stats_lock,
+            )
+
+            # Create a completed future for the task tracking system
+            node_exec_future = Future()
+            node_exec_progress.add_task(
+                node_exec_id=node_exec_result.node_exec_id,
+                task=node_exec_future,
+            )
+
+            # Execute the node directly since we're in the SmartAgent context
+            node_exec_future.set_result(
+                await execution_processor.on_node_execution(
+                    node_exec=node_exec_entry,
+                    node_exec_progress=node_exec_progress,
+                    nodes_input_masks=None,
+                    graph_stats_pair=graph_stats_pair,
+                )
+            )
+
+            # Get outputs from database after execution completes using database manager client
+            node_outputs = await db_client.get_execution_outputs_by_node_exec_id(
+                node_exec_result.node_exec_id
+            )
+
+            # Create tool response
+            tool_response_content = (
+                json.dumps(node_outputs)
+                if node_outputs
+                else "Tool executed successfully"
+            )
+            return _create_tool_response(tool_call.id, tool_response_content)
+        except Exception as e:
+            logger.error(f"Tool execution with manager failed: {e}")
+            # Return error response
+            return _create_tool_response(
+                tool_call.id, f"Tool execution failed: {str(e)}"
+            )
+
+    def _setup_environment(
+        self, credentials: APIKeyCredentials, working_dir: str
+    ) -> dict[str, str]:
+        """Setup environment for Claude Agent SDK."""
+        os.makedirs(working_dir, exist_ok=True)
+        env = os.environ.copy()
+        env["ANTHROPIC_API_KEY"] = credentials.api_key.get_secret_value()
+        return env
+
+    def _build_tool_descriptions_for_claude(
+        self, tool_functions: list[dict[str, Any]]
+    ) -> str:
+        """Build description of available AutoGPT tools for Claude."""
+        if not tool_functions:
+            return "No tools are currently connected to this agent."
+
+        tool_descriptions = ["Available AutoGPT tools:"]
+        for tool_def in tool_functions:
+            func_def = tool_def.get("function", {})
+            name = func_def.get("name", "unknown")
+            description = func_def.get("description", "No description")
+            tool_descriptions.append(f"- {name}: {description}")
+
+        tool_descriptions.append(
+            "\nWhen you need to use a tool, call it with function calling syntax."
+        )
+        return "\n".join(tool_descriptions)
+
+    def _extract_tool_calls_from_claude_message(
+        self, message: AssistantMessage
+    ) -> list[dict[str, Any]]:
+        """Extract tool calls from Claude Agent SDK message."""
+        tool_calls = []
+        for content_block in message.content:
+            if isinstance(content_block, ToolUseBlock):
+                tool_calls.append(
+                    {
+                        "id": content_block.id,
+                        "function": {
+                            "name": content_block.name,
+                            "arguments": content_block.input,
+                        },
+                    }
+                )
+        return tool_calls
+
+    def _extract_text_content_from_claude_message(
+        self, message: AssistantMessage
+    ) -> str:
+        """Extract text content from Claude Agent SDK message."""
+        text_parts = []
+        for content_block in message.content:
+            if isinstance(content_block, TextBlock):
+                text_parts.append(content_block.text)
+        return "".join(text_parts)
+
+    def _format_conversation_for_claude(self, conversation: list[dict]) -> str:
+        """Format conversation history for Claude Agent SDK."""
+        formatted = []
+        for msg in conversation:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "tool":
+                # Format tool response
+                tool_id = msg.get("tool_call_id", "unknown")
+                formatted.append(f"Tool result ({tool_id}): {content}")
+            else:
+                # Simple format for user/assistant messages
+                formatted.append(f"{role.title()}: {content}")
+
+        return "\n\n".join(formatted)
+
+    def _normalize_tool_args(self, tool_args: Any) -> dict:
+        """Normalize tool arguments to dict format."""
+        if isinstance(tool_args, str):
+            return json.loads(tool_args)
+        elif isinstance(tool_args, dict):
+            return tool_args
+        else:
+            return dict(tool_args) if tool_args else {}
+
+    def _create_tool_info_from_claude_call(
+        self, tool_call: dict[str, Any], tool_functions: list[dict[str, Any]]
+    ) -> ToolInfo:
+        """Convert Claude tool call to AutoGPT ToolInfo format."""
+        tool_name = tool_call["function"]["name"]
+        tool_args = self._normalize_tool_args(tool_call["function"]["arguments"])
+        tool_id = tool_call["id"]
+
+        # Find the AutoGPT tool definition
+        tool_def = next(
+            (
+                tf
+                for tf in tool_functions
+                if tf.get("function", {}).get("name") == tool_name
+            ),
+            None,
+        )
+
+        if not tool_def:
+            raise ValueError(f"AutoGPT tool '{tool_name}' not found")
+
+        # Create mock tool call object for AutoGPT compatibility
+        class MockToolCall:
+            def __init__(self, tool_id: str, name: str, args: dict):
+                self.id = tool_id
+                self.function = type(
+                    "Function", (), {"name": name, "arguments": json.dumps(args)}
+                )()
+
+        # Build input data from arguments
+        field_mapping = tool_def["function"].get("_field_mapping", {})
+        expected_args = tool_def["function"]["parameters"].get("properties", {})
+
+        input_data = {
+            field_mapping.get(clean_arg_name, clean_arg_name): tool_args.get(
+                clean_arg_name
+            )
+            for clean_arg_name in expected_args
+        }
+
+        return ToolInfo(
+            tool_call=MockToolCall(tool_id, tool_name, tool_args),
+            tool_name=tool_name,
+            tool_def=tool_def,
+            input_data=input_data,
+            field_mapping=field_mapping,
+        )
+
+    async def _attempt_claude_call_with_validation(
+        self,
+        prompt: str,
+        options: ClaudeAgentOptions,
+    ) -> AssistantMessage:
+        """Claude SDK call - let generator cleanup happen naturally to avoid cancel scope issues."""
+        try:
+            # Simple approach: don't try to manually manage the generator lifecycle
+            async for message in query(prompt=prompt, options=options):
+                if isinstance(message, AssistantMessage):
+                    return message
+
+            raise ValueError("No AssistantMessage received from Claude SDK")
+
+        except Exception as e:
+            logger.error(f"Claude SDK call failed: {e}")
+            raise
+        # Note: No finally block - let the generator be cleaned up naturally by garbage collection
+
+    async def _execute_tools_agent_mode(
+        self,
+        input_data: Input,
+        credentials,
+        tool_functions: list[dict[str, Any]],
+        graph_exec_id: str,
+        node_id: str,
+        node_exec_id: str,
+        user_id: str,
+        graph_id: str,
+        graph_version: int,
+        execution_context: ExecutionContext,
+        execution_processor: "ExecutionProcessor",
+    ):
+        """Execute tools in agent mode with a loop until finished, following SmartDecisionMakerBlock pattern."""
+        max_iterations = input_data.max_iterations
+        iteration = 0
+
+        # Execution parameters for tool execution
+        execution_params = ExecutionParams(
+            user_id=user_id,
+            graph_id=graph_id,
+            node_id=node_id,
+            graph_version=graph_version,
+            graph_exec_id=graph_exec_id,
+            node_exec_id=node_exec_id,
+            execution_context=execution_context,
+        )
+
+        # Build enhanced task prompt with tool descriptions
+        tool_descriptions = self._build_tool_descriptions_for_claude(tool_functions)
+        enhanced_task = f"""{input_data.task}
+
+{tool_descriptions}
+
+Complete the task step by step using the available tools as needed."""
+
+        # Start conversation with enhanced task
+        current_conversation = [{"role": "user", "content": enhanced_task}]
+
+        while max_iterations < 0 or iteration < max_iterations:
+            iteration += 1
+            logger.debug(f"Claude agent mode iteration {iteration}")
+
+            # Prepare conversation for this iteration
+            iteration_conversation = list(current_conversation)
+
+            # On the last iteration, add encouragement to finish
+            if max_iterations > 0 and iteration == max_iterations:
+                last_iteration_message = {
+                    "role": "system",
+                    "content": f"This is your last iteration ({iteration}/{max_iterations}). "
+                    "Try to complete the task with the information you have. "
+                    "Prefer finishing with a clear response rather than making additional tool calls.",
+                }
+                iteration_conversation.append(last_iteration_message)
+
+            # Format conversation for Claude SDK
+            conversation_text = self._format_conversation_for_claude(
+                iteration_conversation
+            )
+
+            # Setup Claude options for this iteration
+            claude_options = ClaudeAgentOptions(
+                system_prompt=input_data.system_prompt,
+                model=input_data.model.value,
+                max_turns=1,  # Single turn per iteration
+                cwd=input_data.working_directory,
+                env=self._setup_environment(credentials, input_data.working_directory),
+                permission_mode="bypassPermissions",
+            )
+
+            # Get Claude response
+            logger.debug(f"Claude agent iteration {iteration}: Making Claude SDK call")
+            try:
+                claude_response = await self._attempt_claude_call_with_validation(
+                    conversation_text, claude_options
+                )
+                logger.debug(f"Claude agent iteration {iteration}: Received response")
+            except Exception as e:
+                logger.error(
+                    f"Claude agent iteration {iteration}: Call failed with {type(e).__name__}: {str(e)}"
+                )
+                yield (
+                    "error",
+                    f"Claude call failed in agent mode iteration {iteration}: {str(e)}",
+                )
+                return
+
+            # Process tool calls
+            tool_calls = self._extract_tool_calls_from_claude_message(claude_response)
+            text_content = self._extract_text_content_from_claude_message(
+                claude_response
+            )
+
+            # Add Claude's response to conversation
+            assistant_message = {
+                "role": "assistant",
+                "content": text_content,
+                "tool_calls": tool_calls if tool_calls else [],
+            }
+            current_conversation.append(assistant_message)
+
+            # If no tool calls, we're done
+            if not tool_calls:
+                yield "finished", text_content
+                yield "conversations", current_conversation
+                return
+
+            # Execute tools and collect responses
+            tool_outputs = []
+            for tool_call in tool_calls:
+                # Convert tool call to ToolInfo format for AutoGPT execution
+                tool_info = self._create_tool_info_from_claude_call(
+                    tool_call, tool_functions
+                )
+
+                try:
+                    # Execute via AutoGPT's execution manager
+                    tool_response = await self._execute_single_tool_with_manager(
+                        tool_info, execution_params, execution_processor
+                    )
+                    tool_outputs.append(tool_response)
+                except Exception as e:
+                    logger.error(f"Tool execution failed: {e}")
+                    error_response = _create_tool_response(
+                        tool_call["id"], f"Error: {str(e)}"
+                    )
+                    tool_outputs.append(error_response)
+
+            # Add tool results to conversation
+            current_conversation.extend(tool_outputs)
+
+            # Yield intermediate conversation state
+            yield "conversations", current_conversation
+
+        # If we reach max iterations, yield the current state
+        if max_iterations < 0:
+            yield "finished", f"Agent mode completed after {iteration} iterations"
+        else:
+            yield (
+                "finished",
+                f"Agent mode completed after {max_iterations} iterations (limit reached)",
+            )
+        yield "conversations", current_conversation
+
+    async def _execute_single_call_mode(
+        self,
+        input_data: Input,  # Used for configuration and consistency with agent mode
+        tool_functions: list[dict[str, Any]],
+        enhanced_task: str,
+        claude_options: ClaudeAgentOptions | None,
+    ):
+        """Execute single call mode and yield tool outputs for external execution."""
+        # Create Claude options for single call if not provided
+        if claude_options is None:
+            claude_options = ClaudeAgentOptions(
+                system_prompt=input_data.system_prompt,
+                model=input_data.model.value,
+                max_turns=1,  # Single call mode
+                cwd=input_data.working_directory,
+                permission_mode="bypassPermissions",
+            )
+        else:
+            # Override max_turns to 1 for single call
+            claude_options.max_turns = 1
+
+        try:
+            claude_response = await self._attempt_claude_call_with_validation(
+                enhanced_task, claude_options
+            )
+        except Exception as e:
+            yield "error", f"Claude SDK error: {str(e)}"
+            yield "success", False
+            return
+
+        if claude_response:
+            text_content = self._extract_text_content_from_claude_message(
+                claude_response
+            )
+            tool_calls = self._extract_tool_calls_from_claude_message(claude_response)
+
+            if not tool_calls:
+                # No tool calls - just return the result
+                yield "result", text_content
+                yield "success", True
+                yield "tools", []  # No tools used
+                return
+
+            # Process and yield tool calls for external execution
+            for tool_call in tool_calls:
+                tool_name = tool_call["function"]["name"]
+                tool_args = self._normalize_tool_args(
+                    tool_call["function"]["arguments"]
+                )
+
+                # Find the tool definition (fallback to first if only one available)
+                tool_def = next(
+                    (
+                        tool
+                        for tool in tool_functions
+                        if tool["function"]["name"] == tool_name
+                    ),
+                    tool_functions[0] if len(tool_functions) == 1 else None,
+                )
+                if not tool_def:
+                    continue
+
+                # Get field mapping and sink node ID
+                field_mapping = tool_def["function"].get("_field_mapping", {})
+                sink_node_id = tool_def["function"]["_sink_node_id"]
+                expected_args = tool_def["function"]["parameters"].get(
+                    "properties", tool_args.keys()
+                )
+
+                # Yield tool outputs like SmartDecisionMakerBlock
+                for clean_arg_name in expected_args:
+                    original_field_name = field_mapping.get(
+                        clean_arg_name, clean_arg_name
+                    )
+                    arg_value = tool_args.get(clean_arg_name)
+
+                    # Create the same emit key format as SmartDecisionMakerBlock
+                    sanitized_arg_name = self.cleanup(original_field_name)
+                    emit_key = f"tools_^_{sink_node_id}_~_{sanitized_arg_name}"
+
+                    logger.debug(f"Yielding tool output: {emit_key}")
+                    yield emit_key, arg_value
+
+            # Yield conversation and tool results
+            yield (
+                "conversations",
+                [
+                    {
+                        "role": "assistant",
+                        "content": text_content,
+                        "tool_calls": tool_calls,
+                    }
+                ],
+            )
+            yield "tools", tool_calls
+            yield "success", True
+            return
+
+        # If no messages received
+        yield "error", "No response from Claude Agent SDK"
+        yield "success", False
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        credentials: APIKeyCredentials,
+        graph_id: str,
+        node_id: str,
+        graph_exec_id: str,
+        node_exec_id: str,
+        user_id: str,
+        graph_version: int,
+        execution_context: ExecutionContext,
+        execution_processor: "ExecutionProcessor",
+        **kwargs,  # Additional execution context parameters
+    ) -> BlockOutput:
+        _ = kwargs  # Suppress unused parameter warning
+        # Validate credentials
+        if credentials.provider != ProviderName.ANTHROPIC:
+            error_msg = f"SmartAgentBlock requires Anthropic/Claude credentials, but received {credentials.provider} credentials. Please configure Anthropic API key credentials."
+            logger.error(error_msg)
+            yield "error", error_msg
+            yield "success", False
+            return
+        # Discover connected AutoGPT tool nodes
+        try:
+            tool_functions = await self._create_tool_node_signatures(node_id)
+        except ValueError as e:
+            if "no next node" in str(e).lower():
+                # Agent can work without tools - just provide Claude with reasoning capability
+                tool_functions = []
+                logger.info("No tools connected - running as pure Claude Agent")
+            else:
+                raise
+
+        yield "tool_functions", json.dumps(tool_functions)
+
+        # Always run Claude Agent SDK in agent mode (iterative execution)
+        async for result in self._execute_tools_agent_mode(
+            input_data=input_data,
+            credentials=credentials,
+            tool_functions=tool_functions,
+            graph_exec_id=graph_exec_id,
+            node_id=node_id,
+            node_exec_id=node_exec_id,
+            user_id=user_id,
+            graph_id=graph_id,
+            graph_version=graph_version,
+            execution_context=execution_context,
+            execution_processor=execution_processor,
+        ):
+            yield result

autogpt_platform/backend/backend/blocks/smart_decision_maker.py

@@ -1,8 +1,11 @@
 import logging
 import re
 from collections import Counter
+from concurrent.futures import Future
 from typing import TYPE_CHECKING, Any
 
+from pydantic import BaseModel
+
 import backend.blocks.llm as llm
 from backend.blocks.agent import AgentExecutorBlock
 from backend.data.block import (
@@ -20,16 +23,41 @@ from backend.data.dynamic_fields import (
     is_dynamic_field,
     is_tool_pin,
 )
+from backend.data.execution import ExecutionContext
 from backend.data.model import NodeExecutionStats, SchemaField
 from backend.util import json
 from backend.util.clients import get_database_manager_async_client
+from backend.util.prompt import MAIN_OBJECTIVE_PREFIX
 
 if TYPE_CHECKING:
     from backend.data.graph import Link, Node
+    from backend.executor.manager import ExecutionProcessor
 
 logger = logging.getLogger(__name__)
 
 
+class ToolInfo(BaseModel):
+    """Processed tool call information."""
+
+    tool_call: Any  # The original tool call object from LLM response
+    tool_name: str  # The function name
+    tool_def: dict[str, Any]  # The tool definition from tool_functions
+    input_data: dict[str, Any]  # Processed input data ready for tool execution
+    field_mapping: dict[str, str]  # Field name mapping for the tool
+
+
+class ExecutionParams(BaseModel):
+    """Tool execution parameters."""
+
+    user_id: str
+    graph_id: str
+    node_id: str
+    graph_version: int
+    graph_exec_id: str
+    node_exec_id: str
+    execution_context: "ExecutionContext"
+
+
 def _get_tool_requests(entry: dict[str, Any]) -> list[str]:
     """
     Return a list of tool_call_ids if the entry is a tool request.
@@ -105,6 +133,50 @@ def _create_tool_response(call_id: str, output: Any) -> dict[str, Any]:
     return {"role": "tool", "tool_call_id": call_id, "content": content}
 
 
+def _combine_tool_responses(tool_outputs: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """
+    Combine multiple Anthropic tool responses into a single user message.
+    For non-Anthropic formats, returns the original list unchanged.
+    """
+    if len(tool_outputs) <= 1:
+        return tool_outputs
+
+    # Anthropic responses have role="user", type="message", and content is a list with tool_result items
+    anthropic_responses = [
+        output
+        for output in tool_outputs
+        if (
+            output.get("role") == "user"
+            and output.get("type") == "message"
+            and isinstance(output.get("content"), list)
+            and any(
+                item.get("type") == "tool_result"
+                for item in output.get("content", [])
+                if isinstance(item, dict)
+            )
+        )
+    ]
+
+    if len(anthropic_responses) > 1:
+        combined_content = [
+            item for response in anthropic_responses for item in response["content"]
+        ]
+
+        combined_response = {
+            "role": "user",
+            "type": "message",
+            "content": combined_content,
+        }
+
+        non_anthropic_responses = [
+            output for output in tool_outputs if output not in anthropic_responses
+        ]
+
+        return [combined_response] + non_anthropic_responses
+
+    return tool_outputs
+
+
 def _convert_raw_response_to_dict(raw_response: Any) -> dict[str, Any]:
     """
     Safely convert raw_response to dictionary format for conversation history.
@@ -204,6 +276,17 @@ class SmartDecisionMakerBlock(Block):
             default="localhost:11434",
             description="Ollama host for local  models",
         )
+        agent_mode_max_iterations: int = SchemaField(
+            title="Agent Mode Max Iterations",
+            description="Maximum iterations for agent mode. 0 = traditional mode (single LLM call, yield tool calls for external execution), -1 = infinite agent mode (loop until finished), 1+ = agent mode with max iterations limit.",
+            advanced=True,
+            default=0,
+        )
+        conversation_compaction: bool = SchemaField(
+            default=True,
+            title="Context window auto-compaction",
+            description="Automatically compact the context window once it hits the limit",
+        )
 
         @classmethod
         def get_missing_links(cls, data: BlockInput, links: list["Link"]) -> set[str]:
@@ -506,6 +589,7 @@ class SmartDecisionMakerBlock(Block):
         Returns the response if successful, raises ValueError if validation fails.
         """
         resp = await llm.llm_call(
+            compress_prompt_to_fit=input_data.conversation_compaction,
             credentials=credentials,
             llm_model=input_data.model,
             prompt=current_prompt,
@@ -593,6 +677,291 @@ class SmartDecisionMakerBlock(Block):
 
         return resp
 
+    def _process_tool_calls(
+        self, response, tool_functions: list[dict[str, Any]]
+    ) -> list[ToolInfo]:
+        """Process tool calls and extract tool definitions, arguments, and input data.
+
+        Returns a list of tool info dicts with:
+        - tool_call: The original tool call object
+        - tool_name: The function name
+        - tool_def: The tool definition from tool_functions
+        - input_data: Processed input data dict (includes None values)
+        - field_mapping: Field name mapping for the tool
+        """
+        if not response.tool_calls:
+            return []
+
+        processed_tools = []
+        for tool_call in response.tool_calls:
+            tool_name = tool_call.function.name
+            tool_args = json.loads(tool_call.function.arguments)
+
+            tool_def = next(
+                (
+                    tool
+                    for tool in tool_functions
+                    if tool["function"]["name"] == tool_name
+                ),
+                None,
+            )
+            if not tool_def:
+                if len(tool_functions) == 1:
+                    tool_def = tool_functions[0]
+                else:
+                    continue
+
+            # Build input data for the tool
+            input_data = {}
+            field_mapping = tool_def["function"].get("_field_mapping", {})
+            if "function" in tool_def and "parameters" in tool_def["function"]:
+                expected_args = tool_def["function"]["parameters"].get("properties", {})
+                for clean_arg_name in expected_args:
+                    original_field_name = field_mapping.get(
+                        clean_arg_name, clean_arg_name
+                    )
+                    arg_value = tool_args.get(clean_arg_name)
+                    # Include all expected parameters, even if None (for backward compatibility with tests)
+                    input_data[original_field_name] = arg_value
+
+            processed_tools.append(
+                ToolInfo(
+                    tool_call=tool_call,
+                    tool_name=tool_name,
+                    tool_def=tool_def,
+                    input_data=input_data,
+                    field_mapping=field_mapping,
+                )
+            )
+
+        return processed_tools
+
+    def _update_conversation(
+        self, prompt: list[dict], response, tool_outputs: list | None = None
+    ):
+        """Update conversation history with response and tool outputs."""
+        # Don't add separate reasoning message with tool calls (breaks Anthropic's tool_use->tool_result pairing)
+        assistant_message = _convert_raw_response_to_dict(response.raw_response)
+        has_tool_calls = isinstance(assistant_message.get("content"), list) and any(
+            item.get("type") == "tool_use"
+            for item in assistant_message.get("content", [])
+        )
+
+        if response.reasoning and not has_tool_calls:
+            prompt.append(
+                {"role": "assistant", "content": f"[Reasoning]: {response.reasoning}"}
+            )
+
+        prompt.append(assistant_message)
+
+        if tool_outputs:
+            prompt.extend(tool_outputs)
+
+    async def _execute_single_tool_with_manager(
+        self,
+        tool_info: ToolInfo,
+        execution_params: ExecutionParams,
+        execution_processor: "ExecutionProcessor",
+    ) -> dict:
+        """Execute a single tool using the execution manager for proper integration."""
+        # Lazy imports to avoid circular dependencies
+        from backend.data.execution import NodeExecutionEntry
+
+        tool_call = tool_info.tool_call
+        tool_def = tool_info.tool_def
+        raw_input_data = tool_info.input_data
+
+        # Get sink node and field mapping
+        sink_node_id = tool_def["function"]["_sink_node_id"]
+
+        # Use proper database operations for tool execution
+        db_client = get_database_manager_async_client()
+
+        # Get target node
+        target_node = await db_client.get_node(sink_node_id)
+        if not target_node:
+            raise ValueError(f"Target node {sink_node_id} not found")
+
+        # Create proper node execution using upsert_execution_input
+        node_exec_result = None
+        final_input_data = None
+
+        # Add all inputs to the execution
+        if not raw_input_data:
+            raise ValueError(f"Tool call has no input data: {tool_call}")
+
+        for input_name, input_value in raw_input_data.items():
+            node_exec_result, final_input_data = await db_client.upsert_execution_input(
+                node_id=sink_node_id,
+                graph_exec_id=execution_params.graph_exec_id,
+                input_name=input_name,
+                input_data=input_value,
+            )
+
+        assert node_exec_result is not None, "node_exec_result should not be None"
+
+        # Create NodeExecutionEntry for execution manager
+        node_exec_entry = NodeExecutionEntry(
+            user_id=execution_params.user_id,
+            graph_exec_id=execution_params.graph_exec_id,
+            graph_id=execution_params.graph_id,
+            graph_version=execution_params.graph_version,
+            node_exec_id=node_exec_result.node_exec_id,
+            node_id=sink_node_id,
+            block_id=target_node.block_id,
+            inputs=final_input_data or {},
+            execution_context=execution_params.execution_context,
+        )
+
+        # Use the execution manager to execute the tool node
+        try:
+            # Get NodeExecutionProgress from the execution manager's running nodes
+            node_exec_progress = execution_processor.running_node_execution[
+                sink_node_id
+            ]
+
+            # Use the execution manager's own graph stats
+            graph_stats_pair = (
+                execution_processor.execution_stats,
+                execution_processor.execution_stats_lock,
+            )
+
+            # Create a completed future for the task tracking system
+            node_exec_future = Future()
+            node_exec_progress.add_task(
+                node_exec_id=node_exec_result.node_exec_id,
+                task=node_exec_future,
+            )
+
+            # Execute the node directly since we're in the SmartDecisionMaker context
+            node_exec_future.set_result(
+                await execution_processor.on_node_execution(
+                    node_exec=node_exec_entry,
+                    node_exec_progress=node_exec_progress,
+                    nodes_input_masks=None,
+                    graph_stats_pair=graph_stats_pair,
+                )
+            )
+
+            # Get outputs from database after execution completes using database manager client
+            node_outputs = await db_client.get_execution_outputs_by_node_exec_id(
+                node_exec_result.node_exec_id
+            )
+
+            # Create tool response
+            tool_response_content = (
+                json.dumps(node_outputs)
+                if node_outputs
+                else "Tool executed successfully"
+            )
+            return _create_tool_response(tool_call.id, tool_response_content)
+
+        except Exception as e:
+            logger.error(f"Tool execution with manager failed: {e}")
+            # Return error response
+            return _create_tool_response(
+                tool_call.id, f"Tool execution failed: {str(e)}"
+            )
+
+    async def _execute_tools_agent_mode(
+        self,
+        input_data,
+        credentials,
+        tool_functions: list[dict[str, Any]],
+        prompt: list[dict],
+        graph_exec_id: str,
+        node_id: str,
+        node_exec_id: str,
+        user_id: str,
+        graph_id: str,
+        graph_version: int,
+        execution_context: ExecutionContext,
+        execution_processor: "ExecutionProcessor",
+    ):
+        """Execute tools in agent mode with a loop until finished."""
+        max_iterations = input_data.agent_mode_max_iterations
+        iteration = 0
+
+        # Execution parameters for tool execution
+        execution_params = ExecutionParams(
+            user_id=user_id,
+            graph_id=graph_id,
+            node_id=node_id,
+            graph_version=graph_version,
+            graph_exec_id=graph_exec_id,
+            node_exec_id=node_exec_id,
+            execution_context=execution_context,
+        )
+
+        current_prompt = list(prompt)
+
+        while max_iterations < 0 or iteration < max_iterations:
+            iteration += 1
+            logger.debug(f"Agent mode iteration {iteration}")
+
+            # Prepare prompt for this iteration
+            iteration_prompt = list(current_prompt)
+
+            # On the last iteration, add a special system message to encourage completion
+            if max_iterations > 0 and iteration == max_iterations:
+                last_iteration_message = {
+                    "role": "system",
+                    "content": f"{MAIN_OBJECTIVE_PREFIX}This is your last iteration ({iteration}/{max_iterations}). "
+                    "Try to complete the task with the information you have. If you cannot fully complete it, "
+                    "provide a summary of what you've accomplished and what remains to be done. "
+                    "Prefer finishing with a clear response rather than making additional tool calls.",
+                }
+                iteration_prompt.append(last_iteration_message)
+
+            # Get LLM response
+            try:
+                response = await self._attempt_llm_call_with_validation(
+                    credentials, input_data, iteration_prompt, tool_functions
+                )
+            except Exception as e:
+                yield "error", f"LLM call failed in agent mode iteration {iteration}: {str(e)}"
+                return
+
+            # Process tool calls
+            processed_tools = self._process_tool_calls(response, tool_functions)
+
+            # If no tool calls, we're done
+            if not processed_tools:
+                yield "finished", response.response
+                self._update_conversation(current_prompt, response)
+                yield "conversations", current_prompt
+                return
+
+            # Execute tools and collect responses
+            tool_outputs = []
+            for tool_info in processed_tools:
+                try:
+                    tool_response = await self._execute_single_tool_with_manager(
+                        tool_info, execution_params, execution_processor
+                    )
+                    tool_outputs.append(tool_response)
+                except Exception as e:
+                    logger.error(f"Tool execution failed: {e}")
+                    # Create error response for the tool
+                    error_response = _create_tool_response(
+                        tool_info.tool_call.id, f"Error: {str(e)}"
+                    )
+                    tool_outputs.append(error_response)
+
+            tool_outputs = _combine_tool_responses(tool_outputs)
+
+            self._update_conversation(current_prompt, response, tool_outputs)
+
+            # Yield intermediate conversation state
+            yield "conversations", current_prompt
+
+        # If we reach max iterations, yield the current state
+        if max_iterations < 0:
+            yield "finished", f"Agent mode completed after {iteration} iterations"
+        else:
+            yield "finished", f"Agent mode completed after {max_iterations} iterations (limit reached)"
+        yield "conversations", current_prompt
+
     async def run(
         self,
         input_data: Input,
@@ -603,8 +972,12 @@ class SmartDecisionMakerBlock(Block):
         graph_exec_id: str,
         node_exec_id: str,
         user_id: str,
+        graph_version: int,
+        execution_context: ExecutionContext,
+        execution_processor: "ExecutionProcessor",
         **kwargs,
     ) -> BlockOutput:
+
         tool_functions = await self._create_tool_node_signatures(node_id)
         yield "tool_functions", json.dumps(tool_functions)
 
@@ -648,24 +1021,52 @@ class SmartDecisionMakerBlock(Block):
             input_data.prompt = llm.fmt.format_string(input_data.prompt, values)
             input_data.sys_prompt = llm.fmt.format_string(input_data.sys_prompt, values)
 
-        prefix = "[Main Objective Prompt]: "
-
         if input_data.sys_prompt and not any(
-            p["role"] == "system" and p["content"].startswith(prefix) for p in prompt
+            p["role"] == "system" and p["content"].startswith(MAIN_OBJECTIVE_PREFIX)
+            for p in prompt
         ):
-            prompt.append({"role": "system", "content": prefix + input_data.sys_prompt})
+            prompt.append(
+                {
+                    "role": "system",
+                    "content": MAIN_OBJECTIVE_PREFIX + input_data.sys_prompt,
+                }
+            )
 
         if input_data.prompt and not any(
-            p["role"] == "user" and p["content"].startswith(prefix) for p in prompt
+            p["role"] == "user" and p["content"].startswith(MAIN_OBJECTIVE_PREFIX)
+            for p in prompt
         ):
-            prompt.append({"role": "user", "content": prefix + input_data.prompt})
+            prompt.append(
+                {"role": "user", "content": MAIN_OBJECTIVE_PREFIX + input_data.prompt}
+            )
 
+        # Execute tools based on the selected mode
+        if input_data.agent_mode_max_iterations != 0:
+            # In agent mode, execute tools directly in a loop until finished
+            async for result in self._execute_tools_agent_mode(
+                input_data=input_data,
+                credentials=credentials,
+                tool_functions=tool_functions,
+                prompt=prompt,
+                graph_exec_id=graph_exec_id,
+                node_id=node_id,
+                node_exec_id=node_exec_id,
+                user_id=user_id,
+                graph_id=graph_id,
+                graph_version=graph_version,
+                execution_context=execution_context,
+                execution_processor=execution_processor,
+            ):
+                yield result
+            return
+
+        # One-off mode: single LLM call and yield tool calls for external execution
         current_prompt = list(prompt)
         max_attempts = max(1, int(input_data.retry))
         response = None
 
         last_error = None
-        for attempt in range(max_attempts):
+        for _ in range(max_attempts):
             try:
                 response = await self._attempt_llm_call_with_validation(
                     credentials, input_data, current_prompt, tool_functions

autogpt_platform/backend/backend/blocks/test/test_smart_decision_maker.py

@@ -1,7 +1,11 @@
 import logging
+import threading
+from collections import defaultdict
+from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 
+from backend.data.execution import ExecutionContext
 from backend.data.model import ProviderName, User
 from backend.server.model import CreateGraph
 from backend.server.rest_api import AgentServer
@@ -17,10 +21,10 @@ async def create_graph(s: SpinTestServer, g, u: User):
 
 
 async def create_credentials(s: SpinTestServer, u: User):
-    import backend.blocks.llm as llm
+    import backend.blocks.llm as llm_module
 
     provider = ProviderName.OPENAI
-    credentials = llm.TEST_CREDENTIALS
+    credentials = llm_module.TEST_CREDENTIALS
     return await s.agent_server.test_create_credentials(u.id, provider, credentials)
 
 
@@ -196,8 +200,6 @@ async def test_smart_decision_maker_function_signature(server: SpinTestServer):
 @pytest.mark.asyncio
 async def test_smart_decision_maker_tracks_llm_stats():
     """Test that SmartDecisionMakerBlock correctly tracks LLM usage stats."""
-    from unittest.mock import MagicMock, patch
-
     import backend.blocks.llm as llm_module
     from backend.blocks.smart_decision_maker import SmartDecisionMakerBlock
 
@@ -216,7 +218,6 @@ async def test_smart_decision_maker_tracks_llm_stats():
     }
 
     # Mock the _create_tool_node_signatures method to avoid database calls
-    from unittest.mock import AsyncMock
 
     with patch(
         "backend.blocks.llm.llm_call",
@@ -234,10 +235,19 @@ async def test_smart_decision_maker_tracks_llm_stats():
             prompt="Should I continue with this task?",
             model=llm_module.LlmModel.GPT4O,
             credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
+            agent_mode_max_iterations=0,
         )
 
         # Execute the block
         outputs = {}
+        # Create execution context
+
+        mock_execution_context = ExecutionContext(safe_mode=False)
+
+        # Create a mock execution processor for tests
+
+        mock_execution_processor = MagicMock()
+
         async for output_name, output_data in block.run(
             input_data,
             credentials=llm_module.TEST_CREDENTIALS,
@@ -246,6 +256,9 @@ async def test_smart_decision_maker_tracks_llm_stats():
             graph_exec_id="test-exec-id",
             node_exec_id="test-node-exec-id",
             user_id="test-user-id",
+            graph_version=1,
+            execution_context=mock_execution_context,
+            execution_processor=mock_execution_processor,
         ):
             outputs[output_name] = output_data
 
@@ -263,8 +276,6 @@ async def test_smart_decision_maker_tracks_llm_stats():
 @pytest.mark.asyncio
 async def test_smart_decision_maker_parameter_validation():
     """Test that SmartDecisionMakerBlock correctly validates tool call parameters."""
-    from unittest.mock import MagicMock, patch
-
     import backend.blocks.llm as llm_module
     from backend.blocks.smart_decision_maker import SmartDecisionMakerBlock
 
@@ -311,8 +322,6 @@ async def test_smart_decision_maker_parameter_validation():
     mock_response_with_typo.reasoning = None
     mock_response_with_typo.raw_response = {"role": "assistant", "content": None}
 
-    from unittest.mock import AsyncMock
-
     with patch(
         "backend.blocks.llm.llm_call",
         new_callable=AsyncMock,
@@ -329,8 +338,17 @@ async def test_smart_decision_maker_parameter_validation():
             model=llm_module.LlmModel.GPT4O,
             credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
             retry=2,  # Set retry to 2 for testing
+            agent_mode_max_iterations=0,
         )
 
+        # Create execution context
+
+        mock_execution_context = ExecutionContext(safe_mode=False)
+
+        # Create a mock execution processor for tests
+
+        mock_execution_processor = MagicMock()
+
         # Should raise ValueError after retries due to typo'd parameter name
         with pytest.raises(ValueError) as exc_info:
             outputs = {}
@@ -342,6 +360,9 @@ async def test_smart_decision_maker_parameter_validation():
                 graph_exec_id="test-exec-id",
                 node_exec_id="test-node-exec-id",
                 user_id="test-user-id",
+                graph_version=1,
+                execution_context=mock_execution_context,
+                execution_processor=mock_execution_processor,
             ):
                 outputs[output_name] = output_data
 
@@ -368,8 +389,6 @@ async def test_smart_decision_maker_parameter_validation():
     mock_response_missing_required.reasoning = None
     mock_response_missing_required.raw_response = {"role": "assistant", "content": None}
 
-    from unittest.mock import AsyncMock
-
     with patch(
         "backend.blocks.llm.llm_call",
         new_callable=AsyncMock,
@@ -385,8 +404,17 @@ async def test_smart_decision_maker_parameter_validation():
             prompt="Search for keywords",
             model=llm_module.LlmModel.GPT4O,
             credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
+            agent_mode_max_iterations=0,
         )
 
+        # Create execution context
+
+        mock_execution_context = ExecutionContext(safe_mode=False)
+
+        # Create a mock execution processor for tests
+
+        mock_execution_processor = MagicMock()
+
         # Should raise ValueError due to missing required parameter
         with pytest.raises(ValueError) as exc_info:
             outputs = {}
@@ -398,6 +426,9 @@ async def test_smart_decision_maker_parameter_validation():
                 graph_exec_id="test-exec-id",
                 node_exec_id="test-node-exec-id",
                 user_id="test-user-id",
+                graph_version=1,
+                execution_context=mock_execution_context,
+                execution_processor=mock_execution_processor,
             ):
                 outputs[output_name] = output_data
 
@@ -418,8 +449,6 @@ async def test_smart_decision_maker_parameter_validation():
     mock_response_valid.reasoning = None
     mock_response_valid.raw_response = {"role": "assistant", "content": None}
 
-    from unittest.mock import AsyncMock
-
     with patch(
         "backend.blocks.llm.llm_call",
         new_callable=AsyncMock,
@@ -435,10 +464,19 @@ async def test_smart_decision_maker_parameter_validation():
             prompt="Search for keywords",
             model=llm_module.LlmModel.GPT4O,
             credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
+            agent_mode_max_iterations=0,
         )
 
         # Should succeed - optional parameter missing is OK
         outputs = {}
+        # Create execution context
+
+        mock_execution_context = ExecutionContext(safe_mode=False)
+
+        # Create a mock execution processor for tests
+
+        mock_execution_processor = MagicMock()
+
         async for output_name, output_data in block.run(
             input_data,
             credentials=llm_module.TEST_CREDENTIALS,
@@ -447,6 +485,9 @@ async def test_smart_decision_maker_parameter_validation():
             graph_exec_id="test-exec-id",
             node_exec_id="test-node-exec-id",
             user_id="test-user-id",
+            graph_version=1,
+            execution_context=mock_execution_context,
+            execution_processor=mock_execution_processor,
         ):
             outputs[output_name] = output_data
 
@@ -472,8 +513,6 @@ async def test_smart_decision_maker_parameter_validation():
     mock_response_all_params.reasoning = None
     mock_response_all_params.raw_response = {"role": "assistant", "content": None}
 
-    from unittest.mock import AsyncMock
-
     with patch(
         "backend.blocks.llm.llm_call",
         new_callable=AsyncMock,
@@ -489,10 +528,19 @@ async def test_smart_decision_maker_parameter_validation():
             prompt="Search for keywords",
             model=llm_module.LlmModel.GPT4O,
             credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
+            agent_mode_max_iterations=0,
         )
 
         # Should succeed with all parameters
         outputs = {}
+        # Create execution context
+
+        mock_execution_context = ExecutionContext(safe_mode=False)
+
+        # Create a mock execution processor for tests
+
+        mock_execution_processor = MagicMock()
+
         async for output_name, output_data in block.run(
             input_data,
             credentials=llm_module.TEST_CREDENTIALS,
@@ -501,6 +549,9 @@ async def test_smart_decision_maker_parameter_validation():
             graph_exec_id="test-exec-id",
             node_exec_id="test-node-exec-id",
             user_id="test-user-id",
+            graph_version=1,
+            execution_context=mock_execution_context,
+            execution_processor=mock_execution_processor,
         ):
             outputs[output_name] = output_data
 
@@ -513,8 +564,6 @@ async def test_smart_decision_maker_parameter_validation():
 @pytest.mark.asyncio
 async def test_smart_decision_maker_raw_response_conversion():
     """Test that SmartDecisionMaker correctly handles different raw_response types with retry mechanism."""
-    from unittest.mock import MagicMock, patch
-
     import backend.blocks.llm as llm_module
     from backend.blocks.smart_decision_maker import SmartDecisionMakerBlock
 
@@ -584,7 +633,6 @@ async def test_smart_decision_maker_raw_response_conversion():
     )
 
     # Mock llm_call to return different responses on different calls
-    from unittest.mock import AsyncMock
 
     with patch(
         "backend.blocks.llm.llm_call", new_callable=AsyncMock
@@ -603,10 +651,19 @@ async def test_smart_decision_maker_raw_response_conversion():
             model=llm_module.LlmModel.GPT4O,
             credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
             retry=2,
+            agent_mode_max_iterations=0,
         )
 
         # Should succeed after retry, demonstrating our helper function works
         outputs = {}
+        # Create execution context
+
+        mock_execution_context = ExecutionContext(safe_mode=False)
+
+        # Create a mock execution processor for tests
+
+        mock_execution_processor = MagicMock()
+
         async for output_name, output_data in block.run(
             input_data,
             credentials=llm_module.TEST_CREDENTIALS,
@@ -615,6 +672,9 @@ async def test_smart_decision_maker_raw_response_conversion():
             graph_exec_id="test-exec-id",
             node_exec_id="test-node-exec-id",
             user_id="test-user-id",
+            graph_version=1,
+            execution_context=mock_execution_context,
+            execution_processor=mock_execution_processor,
         ):
             outputs[output_name] = output_data
 
@@ -650,8 +710,6 @@ async def test_smart_decision_maker_raw_response_conversion():
         "I'll help you with that."  # Ollama returns string
     )
 
-    from unittest.mock import AsyncMock
-
     with patch(
         "backend.blocks.llm.llm_call",
         new_callable=AsyncMock,
@@ -666,9 +724,18 @@ async def test_smart_decision_maker_raw_response_conversion():
             prompt="Simple prompt",
             model=llm_module.LlmModel.GPT4O,
             credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
+            agent_mode_max_iterations=0,
         )
 
         outputs = {}
+        # Create execution context
+
+        mock_execution_context = ExecutionContext(safe_mode=False)
+
+        # Create a mock execution processor for tests
+
+        mock_execution_processor = MagicMock()
+
         async for output_name, output_data in block.run(
             input_data,
             credentials=llm_module.TEST_CREDENTIALS,
@@ -677,6 +744,9 @@ async def test_smart_decision_maker_raw_response_conversion():
             graph_exec_id="test-exec-id",
             node_exec_id="test-node-exec-id",
             user_id="test-user-id",
+            graph_version=1,
+            execution_context=mock_execution_context,
+            execution_processor=mock_execution_processor,
         ):
             outputs[output_name] = output_data
 
@@ -696,8 +766,6 @@ async def test_smart_decision_maker_raw_response_conversion():
         "content": "Test response",
     }  # Dict format
 
-    from unittest.mock import AsyncMock
-
     with patch(
         "backend.blocks.llm.llm_call",
         new_callable=AsyncMock,
@@ -712,6 +780,160 @@ async def test_smart_decision_maker_raw_response_conversion():
             prompt="Another test",
             model=llm_module.LlmModel.GPT4O,
             credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
+            agent_mode_max_iterations=0,
+        )
+
+        outputs = {}
+        # Create execution context
+
+        mock_execution_context = ExecutionContext(safe_mode=False)
+
+        # Create a mock execution processor for tests
+
+        mock_execution_processor = MagicMock()
+
+        async for output_name, output_data in block.run(
+            input_data,
+            credentials=llm_module.TEST_CREDENTIALS,
+            graph_id="test-graph-id",
+            node_id="test-node-id",
+            graph_exec_id="test-exec-id",
+            node_exec_id="test-node-exec-id",
+            user_id="test-user-id",
+            graph_version=1,
+            execution_context=mock_execution_context,
+            execution_processor=mock_execution_processor,
+        ):
+            outputs[output_name] = output_data
+
+        assert "finished" in outputs
+        assert outputs["finished"] == "Test response"
+
+
+@pytest.mark.asyncio
+async def test_smart_decision_maker_agent_mode():
+    """Test that agent mode executes tools directly and loops until finished."""
+    import backend.blocks.llm as llm_module
+    from backend.blocks.smart_decision_maker import SmartDecisionMakerBlock
+
+    block = SmartDecisionMakerBlock()
+
+    # Mock tool call that requires multiple iterations
+    mock_tool_call_1 = MagicMock()
+    mock_tool_call_1.id = "call_1"
+    mock_tool_call_1.function.name = "search_keywords"
+    mock_tool_call_1.function.arguments = (
+        '{"query": "test", "max_keyword_difficulty": 50}'
+    )
+
+    mock_response_1 = MagicMock()
+    mock_response_1.response = None
+    mock_response_1.tool_calls = [mock_tool_call_1]
+    mock_response_1.prompt_tokens = 50
+    mock_response_1.completion_tokens = 25
+    mock_response_1.reasoning = "Using search tool"
+    mock_response_1.raw_response = {
+        "role": "assistant",
+        "content": None,
+        "tool_calls": [{"id": "call_1", "type": "function"}],
+    }
+
+    # Final response with no tool calls (finished)
+    mock_response_2 = MagicMock()
+    mock_response_2.response = "Task completed successfully"
+    mock_response_2.tool_calls = []
+    mock_response_2.prompt_tokens = 30
+    mock_response_2.completion_tokens = 15
+    mock_response_2.reasoning = None
+    mock_response_2.raw_response = {
+        "role": "assistant",
+        "content": "Task completed successfully",
+    }
+
+    # Mock the LLM call to return different responses on each iteration
+    llm_call_mock = AsyncMock()
+    llm_call_mock.side_effect = [mock_response_1, mock_response_2]
+
+    # Mock tool node signatures
+    mock_tool_signatures = [
+        {
+            "type": "function",
+            "function": {
+                "name": "search_keywords",
+                "_sink_node_id": "test-sink-node-id",
+                "_field_mapping": {},
+                "parameters": {
+                    "properties": {
+                        "query": {"type": "string"},
+                        "max_keyword_difficulty": {"type": "integer"},
+                    },
+                    "required": ["query", "max_keyword_difficulty"],
+                },
+            },
+        }
+    ]
+
+    # Mock database and execution components
+    mock_db_client = AsyncMock()
+    mock_node = MagicMock()
+    mock_node.block_id = "test-block-id"
+    mock_db_client.get_node.return_value = mock_node
+
+    # Mock upsert_execution_input to return proper NodeExecutionResult and input data
+    mock_node_exec_result = MagicMock()
+    mock_node_exec_result.node_exec_id = "test-tool-exec-id"
+    mock_input_data = {"query": "test", "max_keyword_difficulty": 50}
+    mock_db_client.upsert_execution_input.return_value = (
+        mock_node_exec_result,
+        mock_input_data,
+    )
+
+    # No longer need mock_execute_node since we use execution_processor.on_node_execution
+
+    with patch("backend.blocks.llm.llm_call", llm_call_mock), patch.object(
+        block, "_create_tool_node_signatures", return_value=mock_tool_signatures
+    ), patch(
+        "backend.blocks.smart_decision_maker.get_database_manager_async_client",
+        return_value=mock_db_client,
+    ), patch(
+        "backend.executor.manager.async_update_node_execution_status",
+        new_callable=AsyncMock,
+    ), patch(
+        "backend.integrations.creds_manager.IntegrationCredentialsManager"
+    ):
+
+        # Create a mock execution context
+
+        mock_execution_context = ExecutionContext(
+            safe_mode=False,
+        )
+
+        # Create a mock execution processor for agent mode tests
+
+        mock_execution_processor = AsyncMock()
+        # Configure the execution processor mock with required attributes
+        mock_execution_processor.running_node_execution = defaultdict(MagicMock)
+        mock_execution_processor.execution_stats = MagicMock()
+        mock_execution_processor.execution_stats_lock = threading.Lock()
+
+        # Mock the on_node_execution method to return successful stats
+        mock_node_stats = MagicMock()
+        mock_node_stats.error = None  # No error
+        mock_execution_processor.on_node_execution = AsyncMock(
+            return_value=mock_node_stats
+        )
+
+        # Mock the get_execution_outputs_by_node_exec_id method
+        mock_db_client.get_execution_outputs_by_node_exec_id.return_value = {
+            "result": {"status": "success", "data": "search completed"}
+        }
+
+        # Test agent mode with max_iterations = 3
+        input_data = SmartDecisionMakerBlock.Input(
+            prompt="Complete this task using tools",
+            model=llm_module.LlmModel.GPT4O,
+            credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
+            agent_mode_max_iterations=3,  # Enable agent mode with 3 max iterations
         )
 
         outputs = {}
@@ -723,8 +945,115 @@ async def test_smart_decision_maker_raw_response_conversion():
             graph_exec_id="test-exec-id",
             node_exec_id="test-node-exec-id",
             user_id="test-user-id",
+            graph_version=1,
+            execution_context=mock_execution_context,
+            execution_processor=mock_execution_processor,
         ):
             outputs[output_name] = output_data
 
+        # Verify agent mode behavior
+        assert "tool_functions" in outputs  # tool_functions is yielded in both modes
         assert "finished" in outputs
-        assert outputs["finished"] == "Test response"
+        assert outputs["finished"] == "Task completed successfully"
+        assert "conversations" in outputs
+
+        # Verify the conversation includes tool responses
+        conversations = outputs["conversations"]
+        assert len(conversations) > 2  # Should have multiple conversation entries
+
+        # Verify LLM was called twice (once for tool call, once for finish)
+        assert llm_call_mock.call_count == 2
+
+        # Verify tool was executed via execution processor
+        assert mock_execution_processor.on_node_execution.call_count == 1
+
+
+@pytest.mark.asyncio
+async def test_smart_decision_maker_traditional_mode_default():
+    """Test that default behavior (agent_mode_max_iterations=0) works as traditional mode."""
+    import backend.blocks.llm as llm_module
+    from backend.blocks.smart_decision_maker import SmartDecisionMakerBlock
+
+    block = SmartDecisionMakerBlock()
+
+    # Mock tool call
+    mock_tool_call = MagicMock()
+    mock_tool_call.function.name = "search_keywords"
+    mock_tool_call.function.arguments = (
+        '{"query": "test", "max_keyword_difficulty": 50}'
+    )
+
+    mock_response = MagicMock()
+    mock_response.response = None
+    mock_response.tool_calls = [mock_tool_call]
+    mock_response.prompt_tokens = 50
+    mock_response.completion_tokens = 25
+    mock_response.reasoning = None
+    mock_response.raw_response = {"role": "assistant", "content": None}
+
+    mock_tool_signatures = [
+        {
+            "type": "function",
+            "function": {
+                "name": "search_keywords",
+                "_sink_node_id": "test-sink-node-id",
+                "_field_mapping": {},
+                "parameters": {
+                    "properties": {
+                        "query": {"type": "string"},
+                        "max_keyword_difficulty": {"type": "integer"},
+                    },
+                    "required": ["query", "max_keyword_difficulty"],
+                },
+            },
+        }
+    ]
+
+    with patch(
+        "backend.blocks.llm.llm_call",
+        new_callable=AsyncMock,
+        return_value=mock_response,
+    ), patch.object(
+        block, "_create_tool_node_signatures", return_value=mock_tool_signatures
+    ):
+
+        # Test default behavior (traditional mode)
+        input_data = SmartDecisionMakerBlock.Input(
+            prompt="Test prompt",
+            model=llm_module.LlmModel.GPT4O,
+            credentials=llm_module.TEST_CREDENTIALS_INPUT,  # type: ignore
+            agent_mode_max_iterations=0,  # Traditional mode
+        )
+
+        # Create execution context
+
+        mock_execution_context = ExecutionContext(safe_mode=False)
+
+        # Create a mock execution processor for tests
+
+        mock_execution_processor = MagicMock()
+
+        outputs = {}
+        async for output_name, output_data in block.run(
+            input_data,
+            credentials=llm_module.TEST_CREDENTIALS,
+            graph_id="test-graph-id",
+            node_id="test-node-id",
+            graph_exec_id="test-exec-id",
+            node_exec_id="test-node-exec-id",
+            user_id="test-user-id",
+            graph_version=1,
+            execution_context=mock_execution_context,
+            execution_processor=mock_execution_processor,
+        ):
+            outputs[output_name] = output_data
+
+        # Verify traditional mode behavior
+        assert (
+            "tool_functions" in outputs
+        )  # Should yield tool_functions in traditional mode
+        assert (
+            "tools_^_test-sink-node-id_~_query" in outputs
+        )  # Should yield individual tool parameters
+        assert "tools_^_test-sink-node-id_~_max_keyword_difficulty" in outputs
+        assert "conversations" in outputs

autogpt_platform/backend/backend/blocks/test/test_smart_decision_maker_dynamic_fields.py

@@ -1,7 +1,7 @@
 """Comprehensive tests for SmartDecisionMakerBlock dynamic field handling."""
 
 import json
-from unittest.mock import AsyncMock, Mock, patch
+from unittest.mock import AsyncMock, MagicMock, Mock, patch
 
 import pytest
 
@@ -308,10 +308,47 @@ async def test_output_yielding_with_dynamic_fields():
     ) as mock_llm:
         mock_llm.return_value = mock_response
 
-        # Mock the function signature creation
-        with patch.object(
+        # Mock the database manager to avoid HTTP calls during tool execution
+        with patch(
+            "backend.blocks.smart_decision_maker.get_database_manager_async_client"
+        ) as mock_db_manager, patch.object(
             block, "_create_tool_node_signatures", new_callable=AsyncMock
         ) as mock_sig:
+            # Set up the mock database manager
+            mock_db_client = AsyncMock()
+            mock_db_manager.return_value = mock_db_client
+
+            # Mock the node retrieval
+            mock_target_node = Mock()
+            mock_target_node.id = "test-sink-node-id"
+            mock_target_node.block_id = "CreateDictionaryBlock"
+            mock_target_node.block = Mock()
+            mock_target_node.block.name = "Create Dictionary"
+            mock_db_client.get_node.return_value = mock_target_node
+
+            # Mock the execution result creation
+            mock_node_exec_result = Mock()
+            mock_node_exec_result.node_exec_id = "mock-node-exec-id"
+            mock_final_input_data = {
+                "values_#_name": "Alice",
+                "values_#_age": 30,
+                "values_#_email": "alice@example.com",
+            }
+            mock_db_client.upsert_execution_input.return_value = (
+                mock_node_exec_result,
+                mock_final_input_data,
+            )
+
+            # Mock the output retrieval
+            mock_outputs = {
+                "values_#_name": "Alice",
+                "values_#_age": 30,
+                "values_#_email": "alice@example.com",
+            }
+            mock_db_client.get_execution_outputs_by_node_exec_id.return_value = (
+                mock_outputs
+            )
+
             mock_sig.return_value = [
                 {
                     "type": "function",
@@ -337,10 +374,16 @@ async def test_output_yielding_with_dynamic_fields():
                 prompt="Create a user dictionary",
                 credentials=llm.TEST_CREDENTIALS_INPUT,
                 model=llm.LlmModel.GPT4O,
+                agent_mode_max_iterations=0,  # Use traditional mode to test output yielding
             )
 
             # Run the block
             outputs = {}
+            from backend.data.execution import ExecutionContext
+
+            mock_execution_context = ExecutionContext(safe_mode=False)
+            mock_execution_processor = MagicMock()
+
             async for output_name, output_value in block.run(
                 input_data,
                 credentials=llm.TEST_CREDENTIALS,
@@ -349,6 +392,9 @@ async def test_output_yielding_with_dynamic_fields():
                 graph_exec_id="test_exec",
                 node_exec_id="test_node_exec",
                 user_id="test_user",
+                graph_version=1,
+                execution_context=mock_execution_context,
+                execution_processor=mock_execution_processor,
             ):
                 outputs[output_name] = output_value
 
@@ -511,45 +557,108 @@ async def test_validation_errors_dont_pollute_conversation():
                 }
             ]
 
-            # Create input data
-            from backend.blocks import llm
+            # Mock the database manager to avoid HTTP calls during tool execution
+            with patch(
+                "backend.blocks.smart_decision_maker.get_database_manager_async_client"
+            ) as mock_db_manager:
+                # Set up the mock database manager for agent mode
+                mock_db_client = AsyncMock()
+                mock_db_manager.return_value = mock_db_client
 
-            input_data = block.input_schema(
-                prompt="Test prompt",
-                credentials=llm.TEST_CREDENTIALS_INPUT,
-                model=llm.LlmModel.GPT4O,
-                retry=3,  # Allow retries
-            )
+                # Mock the node retrieval
+                mock_target_node = Mock()
+                mock_target_node.id = "test-sink-node-id"
+                mock_target_node.block_id = "TestBlock"
+                mock_target_node.block = Mock()
+                mock_target_node.block.name = "Test Block"
+                mock_db_client.get_node.return_value = mock_target_node
 
-            # Run the block
-            outputs = {}
-            async for output_name, output_value in block.run(
-                input_data,
-                credentials=llm.TEST_CREDENTIALS,
-                graph_id="test_graph",
-                node_id="test_node",
-                graph_exec_id="test_exec",
-                node_exec_id="test_node_exec",
-                user_id="test_user",
-            ):
-                outputs[output_name] = output_value
+                # Mock the execution result creation
+                mock_node_exec_result = Mock()
+                mock_node_exec_result.node_exec_id = "mock-node-exec-id"
+                mock_final_input_data = {"correct_param": "value"}
+                mock_db_client.upsert_execution_input.return_value = (
+                    mock_node_exec_result,
+                    mock_final_input_data,
+                )
 
-            # Verify we had 2 LLM calls (initial + retry)
-            assert call_count == 2
+                # Mock the output retrieval
+                mock_outputs = {"correct_param": "value"}
+                mock_db_client.get_execution_outputs_by_node_exec_id.return_value = (
+                    mock_outputs
+                )
 
-            # Check the final conversation output
-            final_conversation = outputs.get("conversations", [])
+                # Create input data
+                from backend.blocks import llm
 
-            # The final conversation should NOT contain the validation error message
-            error_messages = [
-                msg
-                for msg in final_conversation
-                if msg.get("role") == "user"
-                and "parameter errors" in msg.get("content", "")
-            ]
-            assert (
-                len(error_messages) == 0
-            ), "Validation error leaked into final conversation"
+                input_data = block.input_schema(
+                    prompt="Test prompt",
+                    credentials=llm.TEST_CREDENTIALS_INPUT,
+                    model=llm.LlmModel.GPT4O,
+                    retry=3,  # Allow retries
+                    agent_mode_max_iterations=1,
+                )
 
-            # The final conversation should only have the successful response
-            assert final_conversation[-1]["content"] == "valid"
+                # Run the block
+                outputs = {}
+                from backend.data.execution import ExecutionContext
+
+                mock_execution_context = ExecutionContext(safe_mode=False)
+
+                # Create a proper mock execution processor for agent mode
+                from collections import defaultdict
+
+                mock_execution_processor = AsyncMock()
+                mock_execution_processor.execution_stats = MagicMock()
+                mock_execution_processor.execution_stats_lock = MagicMock()
+
+                # Create a mock NodeExecutionProgress for the sink node
+                mock_node_exec_progress = MagicMock()
+                mock_node_exec_progress.add_task = MagicMock()
+                mock_node_exec_progress.pop_output = MagicMock(
+                    return_value=None
+                )  # No outputs to process
+
+                # Set up running_node_execution as a defaultdict that returns our mock for any key
+                mock_execution_processor.running_node_execution = defaultdict(
+                    lambda: mock_node_exec_progress
+                )
+
+                # Mock the on_node_execution method that gets called during tool execution
+                mock_node_stats = MagicMock()
+                mock_node_stats.error = None
+                mock_execution_processor.on_node_execution.return_value = (
+                    mock_node_stats
+                )
+
+                async for output_name, output_value in block.run(
+                    input_data,
+                    credentials=llm.TEST_CREDENTIALS,
+                    graph_id="test_graph",
+                    node_id="test_node",
+                    graph_exec_id="test_exec",
+                    node_exec_id="test_node_exec",
+                    user_id="test_user",
+                    graph_version=1,
+                    execution_context=mock_execution_context,
+                    execution_processor=mock_execution_processor,
+                ):
+                    outputs[output_name] = output_value
+
+                # Verify we had at least 1 LLM call
+                assert call_count >= 1
+
+                # Check the final conversation output
+                final_conversation = outputs.get("conversations", [])
+
+                # The final conversation should NOT contain validation error messages
+                # Even if retries don't happen in agent mode, we should not leak errors
+                error_messages = [
+                    msg
+                    for msg in final_conversation
+                    if msg.get("role") == "user"
+                    and "parameter errors" in msg.get("content", "")
+                ]
+                assert (
+                    len(error_messages) == 0
+                ), "Validation error leaked into final conversation"

autogpt_platform/backend/backend/data/execution.py

@@ -5,6 +5,7 @@ from enum import Enum
 from multiprocessing import Manager
 from queue import Empty
 from typing import (
+    TYPE_CHECKING,
     Annotated,
     Any,
     AsyncGenerator,
@@ -65,6 +66,9 @@ from .includes import (
 )
 from .model import CredentialsMetaInput, GraphExecutionStats, NodeExecutionStats
 
+if TYPE_CHECKING:
+    pass
+
 T = TypeVar("T")
 
 logger = logging.getLogger(__name__)
@@ -836,6 +840,30 @@ async def upsert_execution_output(
     await AgentNodeExecutionInputOutput.prisma().create(data=data)
 
 
+async def get_execution_outputs_by_node_exec_id(
+    node_exec_id: str,
+) -> dict[str, Any]:
+    """
+    Get all execution outputs for a specific node execution ID.
+
+    Args:
+        node_exec_id: The node execution ID to get outputs for
+
+    Returns:
+        Dictionary mapping output names to their data values
+    """
+    outputs = await AgentNodeExecutionInputOutput.prisma().find_many(
+        where={"referencedByOutputExecId": node_exec_id}
+    )
+
+    result = {}
+    for output in outputs:
+        if output.data is not None:
+            result[output.name] = type_utils.convert(output.data, JsonValue)
+
+    return result
+
+
 async def update_graph_execution_start_time(
     graph_exec_id: str,
 ) -> GraphExecution | None:

autogpt_platform/backend/backend/executor/database.py

@@ -13,6 +13,7 @@ from backend.data.execution import (
     get_block_error_stats,
     get_child_graph_executions,
     get_execution_kv_data,
+    get_execution_outputs_by_node_exec_id,
     get_frequently_executed_graphs,
     get_graph_execution_meta,
     get_graph_executions,
@@ -147,6 +148,7 @@ class DatabaseManager(AppService):
     update_graph_execution_stats = _(update_graph_execution_stats)
     upsert_execution_input = _(upsert_execution_input)
     upsert_execution_output = _(upsert_execution_output)
+    get_execution_outputs_by_node_exec_id = _(get_execution_outputs_by_node_exec_id)
     get_execution_kv_data = _(get_execution_kv_data)
     set_execution_kv_data = _(set_execution_kv_data)
     get_block_error_stats = _(get_block_error_stats)
@@ -277,6 +279,7 @@ class DatabaseManagerAsyncClient(AppServiceClient):
     get_user_integrations = d.get_user_integrations
     upsert_execution_input = d.upsert_execution_input
     upsert_execution_output = d.upsert_execution_output
+    get_execution_outputs_by_node_exec_id = d.get_execution_outputs_by_node_exec_id
     update_graph_execution_stats = d.update_graph_execution_stats
     update_node_execution_status = d.update_node_execution_status
     update_node_execution_status_batch = d.update_node_execution_status_batch

autogpt_platform/backend/backend/executor/manager.py

@@ -133,9 +133,8 @@ def execute_graph(
     cluster_lock: ClusterLock,
 ):
     """Execute graph using thread-local ExecutionProcessor instance"""
-    return _tls.processor.on_graph_execution(
-        graph_exec_entry, cancel_event, cluster_lock
-    )
+    processor: ExecutionProcessor = _tls.processor
+    return processor.on_graph_execution(graph_exec_entry, cancel_event, cluster_lock)
 
 
 T = TypeVar("T")
@@ -143,8 +142,8 @@ T = TypeVar("T")
 
 async def execute_node(
     node: Node,
-    creds_manager: IntegrationCredentialsManager,
     data: NodeExecutionEntry,
+    execution_processor: "ExecutionProcessor",
     execution_stats: NodeExecutionStats | None = None,
     nodes_input_masks: Optional[NodesInputMasks] = None,
 ) -> BlockOutput:
@@ -169,6 +168,7 @@ async def execute_node(
     node_id = data.node_id
     node_block = node.block
     execution_context = data.execution_context
+    creds_manager = execution_processor.creds_manager
 
     log_metadata = LogMetadata(
         logger=_logger,
@@ -212,6 +212,7 @@ async def execute_node(
         "node_exec_id": node_exec_id,
         "user_id": user_id,
         "execution_context": execution_context,
+        "execution_processor": execution_processor,
     }
 
     # Last-minute fetch credentials + acquire a system-wide read-write lock to prevent
@@ -608,8 +609,8 @@ class ExecutionProcessor:
 
             async for output_name, output_data in execute_node(
                 node=node,
-                creds_manager=self.creds_manager,
                 data=node_exec,
+                execution_processor=self,
                 execution_stats=stats,
                 nodes_input_masks=nodes_input_masks,
             ):
@@ -860,12 +861,17 @@ class ExecutionProcessor:
         execution_stats_lock = threading.Lock()
 
         # State holders ----------------------------------------------------
-        running_node_execution: dict[str, NodeExecutionProgress] = defaultdict(
+        self.running_node_execution: dict[str, NodeExecutionProgress] = defaultdict(
             NodeExecutionProgress
         )
-        running_node_evaluation: dict[str, Future] = {}
+        self.running_node_evaluation: dict[str, Future] = {}
+        self.execution_stats = execution_stats
+        self.execution_stats_lock = execution_stats_lock
         execution_queue = ExecutionQueue[NodeExecutionEntry]()
 
+        running_node_execution = self.running_node_execution
+        running_node_evaluation = self.running_node_evaluation
+
         try:
             if db_client.get_credits(graph_exec.user_id) <= 0:
                 raise InsufficientBalanceError(

autogpt_platform/backend/backend/util/prompt.py

@@ -5,6 +5,13 @@ from tiktoken import encoding_for_model
 
 from backend.util import json
 
+# ---------------------------------------------------------------------------#
+#  CONSTANTS                                                                 #
+# ---------------------------------------------------------------------------#
+
+# Message prefixes for important system messages that should be protected during compression
+MAIN_OBJECTIVE_PREFIX = "[Main Objective Prompt]: "
+
 # ---------------------------------------------------------------------------#
 #  INTERNAL UTILITIES                                                         #
 # ---------------------------------------------------------------------------#
@@ -63,6 +70,55 @@ def _msg_tokens(msg: dict, enc) -> int:
     return WRAPPER + content_tokens + tool_call_tokens
 
 
+def _is_tool_message(msg: dict) -> bool:
+    """Check if a message contains tool calls or results that should be protected."""
+    content = msg.get("content")
+
+    # Check for Anthropic-style tool messages
+    if isinstance(content, list) and any(
+        isinstance(item, dict) and item.get("type") in ("tool_use", "tool_result")
+        for item in content
+    ):
+        return True
+
+    # Check for OpenAI-style tool calls in the message
+    if "tool_calls" in msg or msg.get("role") == "tool":
+        return True
+
+    return False
+
+
+def _is_objective_message(msg: dict) -> bool:
+    """Check if a message contains objective/system prompts that should be absolutely protected."""
+    content = msg.get("content", "")
+    if isinstance(content, str):
+        # Protect any message with the main objective prefix
+        return content.startswith(MAIN_OBJECTIVE_PREFIX)
+    return False
+
+
+def _truncate_tool_message_content(msg: dict, enc, max_tokens: int) -> None:
+    """
+    Carefully truncate tool message content while preserving tool structure.
+    Only truncates tool_result content, leaves tool_use intact.
+    """
+    content = msg.get("content")
+    if not isinstance(content, list):
+        return
+
+    for item in content:
+        # Only process tool_result items, leave tool_use blocks completely intact
+        if not (isinstance(item, dict) and item.get("type") == "tool_result"):
+            continue
+
+        result_content = item.get("content", "")
+        if (
+            isinstance(result_content, str)
+            and _tok_len(result_content, enc) > max_tokens
+        ):
+            item["content"] = _truncate_middle_tokens(result_content, enc, max_tokens)
+
+
 def _truncate_middle_tokens(text: str, enc, max_tok: int) -> str:
     """
     Return *text* shortened to ≈max_tok tokens by keeping the head & tail
@@ -140,13 +196,21 @@ def compress_prompt(
         return sum(_msg_tokens(m, enc) for m in msgs)
 
     original_token_count = total_tokens()
+
     if original_token_count + reserve <= target_tokens:
         return msgs
 
     # ---- STEP 0 : normalise content --------------------------------------
     # Convert non-string payloads to strings so token counting is coherent.
-    for m in msgs[1:-1]:  # keep the first & last intact
+    for i, m in enumerate(msgs):
         if not isinstance(m.get("content"), str) and m.get("content") is not None:
+            if _is_tool_message(m):
+                continue
+
+            # Keep first and last messages intact (unless they're tool messages)
+            if i == 0 or i == len(msgs) - 1:
+                continue
+
             # Reasonable 20k-char ceiling prevents pathological blobs
             content_str = json.dumps(m["content"], separators=(",", ":"))
             if len(content_str) > 20_000:
@@ -157,34 +221,45 @@ def compress_prompt(
     cap = start_cap
     while total_tokens() + reserve > target_tokens and cap >= floor_cap:
         for m in msgs[1:-1]:  # keep first & last intact
-            if _tok_len(m.get("content") or "", enc) > cap:
-                m["content"] = _truncate_middle_tokens(m["content"], enc, cap)
+            if _is_tool_message(m):
+                # For tool messages, only truncate tool result content, preserve structure
+                _truncate_tool_message_content(m, enc, cap)
+                continue
+
+            if _is_objective_message(m):
+                # Never truncate objective messages - they contain the core task
+                continue
+
+            content = m.get("content") or ""
+            if _tok_len(content, enc) > cap:
+                m["content"] = _truncate_middle_tokens(content, enc, cap)
         cap //= 2  # tighten the screw
 
     # ---- STEP 2 : middle-out deletion -----------------------------------
     while total_tokens() + reserve > target_tokens and len(msgs) > 2:
+        # Identify all deletable messages (not first/last, not tool messages, not objective messages)
+        deletable_indices = []
+        for i in range(1, len(msgs) - 1):  # Skip first and last
+            if not _is_tool_message(msgs[i]) and not _is_objective_message(msgs[i]):
+                deletable_indices.append(i)
+
+        if not deletable_indices:
+            break  # nothing more we can drop
+
+        # Delete from center outward - find the index closest to center
         centre = len(msgs) // 2
-        # Build a symmetrical centre-out index walk: centre, centre+1, centre-1, ...
-        order = [centre] + [
-            i
-            for pair in zip(range(centre + 1, len(msgs) - 1), range(centre - 1, 0, -1))
-            for i in pair
-        ]
-        removed = False
-        for i in order:
-            msg = msgs[i]
-            if "tool_calls" in msg or msg.get("role") == "tool":
-                continue  # protect tool shells
-            del msgs[i]
-            removed = True
-            break
-        if not removed:  # nothing more we can drop
-            break
+        to_delete = min(deletable_indices, key=lambda i: abs(i - centre))
+        del msgs[to_delete]
 
     # ---- STEP 3 : final safety-net trim on first & last ------------------
     cap = start_cap
     while total_tokens() + reserve > target_tokens and cap >= floor_cap:
         for idx in (0, -1):  # first and last
+            if _is_tool_message(msgs[idx]):
+                # For tool messages at first/last position, truncate tool result content only
+                _truncate_tool_message_content(msgs[idx], enc, cap)
+                continue
+
             text = msgs[idx].get("content") or ""
             if _tok_len(text, enc) > cap:
                 msgs[idx]["content"] = _truncate_middle_tokens(text, enc, cap)

autogpt_platform/backend/poetry.lock

@@ -825,6 +825,29 @@ files = [
     {file = "charset_normalizer-3.4.2.tar.gz", hash = "sha256:5baececa9ecba31eff645232d59845c07aa030f0c81ee70184a90d35099a0e63"},
 ]
 
+[[package]]
+name = "claude-agent-sdk"
+version = "0.1.17"
+description = "Python SDK for Claude Code"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "claude_agent_sdk-0.1.17-py3-none-macosx_11_0_arm64.whl", hash = "sha256:46fd904ce90a183414f979d359519cff347f52c1906d50f1e247551c786a3613"},
+    {file = "claude_agent_sdk-0.1.17-py3-none-manylinux_2_17_aarch64.whl", hash = "sha256:08c41656da3ec7930cb966b35df23da6f68fa0fc0a32c61e73d6450c40c0e17e"},
+    {file = "claude_agent_sdk-0.1.17-py3-none-manylinux_2_17_x86_64.whl", hash = "sha256:be5fe696e50829348b031106dc926bb459c94804a13584a52a483a35ebd1c388"},
+    {file = "claude_agent_sdk-0.1.17-py3-none-win_amd64.whl", hash = "sha256:f2f4e05785a42525cf8026f0bba69bce359b83a57856f0b7349cd311a0d644be"},
+    {file = "claude_agent_sdk-0.1.17.tar.gz", hash = "sha256:85d0818b7259b65ea0ee3103ae736f2b359243de81540ab1eee771d8cae0d8ac"},
+]
+
+[package.dependencies]
+anyio = ">=4.0.0"
+mcp = ">=0.1.0"
+typing-extensions = {version = ">=4.0.0", markers = "python_version < \"3.11\""}
+
+[package.extras]
+dev = ["anyio[trio] (>=4.0.0)", "mypy (>=1.0.0)", "pytest (>=7.0.0)", "pytest-asyncio (>=0.20.0)", "pytest-cov (>=4.0.0)", "ruff (>=0.1.0)"]
+
 [[package]]
 name = "cleo"
 version = "2.1.0"
@@ -2304,6 +2327,18 @@ http2 = ["h2 (>=3,<5)"]
 socks = ["socksio (==1.*)"]
 zstd = ["zstandard (>=0.18.0)"]
 
+[[package]]
+name = "httpx-sse"
+version = "0.4.3"
+description = "Consume Server-Sent Event (SSE) messages with HTTPX."
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "httpx_sse-0.4.3-py3-none-any.whl", hash = "sha256:0ac1c9fe3c0afad2e0ebb25a934a59f4c7823b60792691f779fad2c5568830fc"},
+    {file = "httpx_sse-0.4.3.tar.gz", hash = "sha256:9b1ed0127459a66014aec3c56bebd93da3c1bc8bb6618c8082039a44889a755d"},
+]
+
 [[package]]
 name = "huggingface-hub"
 version = "0.34.4"
@@ -2941,6 +2976,39 @@ files = [
     {file = "mccabe-0.7.0.tar.gz", hash = "sha256:348e0240c33b60bbdf4e523192ef919f28cb2c3d7d5c7794f74009290f236325"},
 ]
 
+[[package]]
+name = "mcp"
+version = "1.24.0"
+description = "Model Context Protocol SDK"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "mcp-1.24.0-py3-none-any.whl", hash = "sha256:db130e103cc50ddc3dffc928382f33ba3eaef0b711f7a87c05e7ded65b1ca062"},
+    {file = "mcp-1.24.0.tar.gz", hash = "sha256:aeaad134664ce56f2721d1abf300666a1e8348563f4d3baff361c3b652448efc"},
+]
+
+[package.dependencies]
+anyio = ">=4.5"
+httpx = ">=0.27.1"
+httpx-sse = ">=0.4"
+jsonschema = ">=4.20.0"
+pydantic = ">=2.11.0,<3.0.0"
+pydantic-settings = ">=2.5.2"
+pyjwt = {version = ">=2.10.1", extras = ["crypto"]}
+python-multipart = ">=0.0.9"
+pywin32 = {version = ">=310", markers = "sys_platform == \"win32\""}
+sse-starlette = ">=1.6.1"
+starlette = ">=0.27"
+typing-extensions = ">=4.9.0"
+typing-inspection = ">=0.4.1"
+uvicorn = {version = ">=0.31.1", markers = "sys_platform != \"emscripten\""}
+
+[package.extras]
+cli = ["python-dotenv (>=1.0.0)", "typer (>=0.16.0)"]
+rich = ["rich (>=13.9.4)"]
+ws = ["websockets (>=15.0.1)"]
+
 [[package]]
 name = "mdurl"
 version = "0.1.2"
@@ -4481,7 +4549,6 @@ files = [
     {file = "psycopg2_binary-2.9.10-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:bb89f0a835bcfc1d42ccd5f41f04870c1b936d8507c6df12b7737febc40f0909"},
     {file = "psycopg2_binary-2.9.10-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:f0c2d907a1e102526dd2986df638343388b94c33860ff3bbe1384130828714b1"},
     {file = "psycopg2_binary-2.9.10-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:f8157bed2f51db683f31306aa497311b560f2265998122abe1dce6428bd86567"},
-    {file = "psycopg2_binary-2.9.10-cp313-cp313-win_amd64.whl", hash = "sha256:27422aa5f11fbcd9b18da48373eb67081243662f9b46e6fd07c3eb46e4535142"},
     {file = "psycopg2_binary-2.9.10-cp38-cp38-macosx_12_0_x86_64.whl", hash = "sha256:eb09aa7f9cecb45027683bb55aebaaf45a0df8bf6de68801a6afdc7947bb09d4"},
     {file = "psycopg2_binary-2.9.10-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b73d6d7f0ccdad7bc43e6d34273f70d587ef62f824d7261c4ae9b8b1b6af90e8"},
     {file = "psycopg2_binary-2.9.10-cp38-cp38-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ce5ab4bf46a211a8e924d307c1b1fcda82368586a19d0a24f8ae166f5c784864"},
@@ -5086,7 +5153,7 @@ description = "Python for Window Extensions"
 optional = false
 python-versions = "*"
 groups = ["main"]
-markers = "platform_system == \"Windows\""
+markers = "platform_system == \"Windows\" or sys_platform == \"win32\""
 files = [
     {file = "pywin32-311-cp310-cp310-win32.whl", hash = "sha256:d03ff496d2a0cd4a5893504789d4a15399133fe82517455e78bad62efbb7f0a3"},
     {file = "pywin32-311-cp310-cp310-win_amd64.whl", hash = "sha256:797c2772017851984b97180b0bebe4b620bb86328e8a884bb626156295a63b3b"},
@@ -6053,6 +6120,27 @@ postgresql-psycopgbinary = ["psycopg[binary] (>=3.0.7)"]
 pymysql = ["pymysql"]
 sqlcipher = ["sqlcipher3_binary"]
 
+[[package]]
+name = "sse-starlette"
+version = "3.0.3"
+description = "SSE plugin for Starlette"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "sse_starlette-3.0.3-py3-none-any.whl", hash = "sha256:af5bf5a6f3933df1d9c7f8539633dc8444ca6a97ab2e2a7cd3b6e431ac03a431"},
+    {file = "sse_starlette-3.0.3.tar.gz", hash = "sha256:88cfb08747e16200ea990c8ca876b03910a23b547ab3bd764c0d8eb81019b971"},
+]
+
+[package.dependencies]
+anyio = ">=4.7.0"
+
+[package.extras]
+daphne = ["daphne (>=4.2.0)"]
+examples = ["aiosqlite (>=0.21.0)", "fastapi (>=0.115.12)", "sqlalchemy[asyncio] (>=2.0.41)", "starlette (>=0.49.1)", "uvicorn (>=0.34.0)"]
+granian = ["granian (>=2.3.1)"]
+uvicorn = ["uvicorn (>=0.34.0)"]
+
 [[package]]
 name = "stagehand"
 version = "0.5.1"
@@ -7279,4 +7367,4 @@ cffi = ["cffi (>=1.11)"]
 [metadata]
 lock-version = "2.1"
 python-versions = ">=3.10,<3.14"
-content-hash = "13b191b2a1989d3321ff713c66ff6f5f4f3b82d15df4d407e0e5dbf87d7522c4"
+content-hash = "4594d56a0d808095e20861774ddfe8f87bc461f4f88ca7b748ff066225fa2684"

autogpt_platform/backend/pyproject.toml

@@ -82,6 +82,7 @@ firecrawl-py = "^4.3.6"
 exa-py = "^1.14.20"
 croniter = "^6.0.0"
 stagehand = "^0.5.1"
+claude-agent-sdk = "^0.1.17"
 
 [tool.poetry.group.dev.dependencies]
 aiohappyeyeballs = "^2.6.1"

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/modals/AgentInputsReadOnly/AgentInputsReadOnly.tsx

@@ -1,16 +1,11 @@
 "use client";
 
 import type { LibraryAgent } from "@/app/api/__generated__/models/libraryAgent";
-import type {
-  BlockIOSubSchema,
-  CredentialsMetaInput,
-} from "@/lib/autogpt-server-api/types";
+import { Text } from "@/components/atoms/Text/Text";
+import type { CredentialsMetaInput } from "@/lib/autogpt-server-api/types";
 import { CredentialsInput } from "../CredentialsInputs/CredentialsInputs";
-import {
-  getAgentCredentialsFields,
-  getAgentInputFields,
-  renderValue,
-} from "./helpers";
+import { RunAgentInputs } from "../RunAgentInputs/RunAgentInputs";
+import { getAgentCredentialsFields, getAgentInputFields } from "./helpers";
 
 type Props = {
   agent: LibraryAgent;
@@ -28,19 +23,23 @@ export function AgentInputsReadOnly({
     getAgentCredentialsFields(agent),
   );
 
-  // Take actual input entries as leading; augment with schema from input fields.
-  // TODO: ensure consistent ordering.
   const inputEntries =
     inputs &&
-    Object.entries(inputs).map<[string, [BlockIOSubSchema | undefined, any]]>(
-      ([k, v]) => [k, [inputFields[k], v]],
-    );
+    Object.entries(inputs).map(([key, value]) => ({
+      key,
+      schema: inputFields[key],
+      value,
+    }));
 
   const hasInputs = inputEntries && inputEntries.length > 0;
   const hasCredentials = credentialInputs && credentialFieldEntries.length > 0;
 
   if (!hasInputs && !hasCredentials) {
-    return <div className="text-neutral-600">No input for this run.</div>;
+    return (
+      <Text variant="body" className="text-zinc-700">
+        No input for this run.
+      </Text>
+    );
   }
 
   return (
@@ -48,16 +47,20 @@ export function AgentInputsReadOnly({
       {/* Regular inputs */}
       {hasInputs && (
         <div className="flex flex-col gap-4">
-          {inputEntries.map(([key, [schema, value]]) => (
-            <div key={key} className="flex flex-col gap-1.5">
-              <label className="text-sm font-medium">
-                {schema?.title || key}
-              </label>
-              <p className="whitespace-pre-wrap break-words text-sm text-neutral-700">
-                {renderValue(value)}
-              </p>
-            </div>
-          ))}
+          {inputEntries.map(({ key, schema, value }) => {
+            if (!schema) return null;
+
+            return (
+              <RunAgentInputs
+                key={key}
+                schema={schema}
+                value={value}
+                placeholder={schema.description}
+                onChange={() => {}}
+                readOnly={true}
+              />
+            );
+          })}
         </div>
       )}
 

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/modals/RunAgentInputs/RunAgentInputs.tsx

@@ -9,6 +9,7 @@ import { Button } from "@/components/atoms/Button/Button";
 import { FileInput } from "@/components/atoms/FileInput/FileInput";
 import { Switch } from "@/components/atoms/Switch/Switch";
 import { GoogleDrivePickerInput } from "@/components/contextual/GoogleDrivePicker/GoogleDrivePickerInput";
+import { InformationTooltip } from "@/components/molecules/InformationTooltip/InformationTooltip";
 import { TimePicker } from "@/components/molecules/TimePicker/TimePicker";
 import {
   BlockIOObjectSubSchema,
@@ -32,6 +33,7 @@ interface Props {
   value?: any;
   placeholder?: string;
   onChange: (value: any) => void;
+  readOnly?: boolean;
 }
 
 /**
@@ -44,6 +46,7 @@ export function RunAgentInputs({
   value,
   placeholder,
   onChange,
+  readOnly = false,
   ...props
 }: Props & React.HTMLAttributes<HTMLElement>) {
   const { handleUploadFile, uploadProgress } = useRunAgentInputs();
@@ -62,7 +65,6 @@ export function RunAgentInputs({
           id={`${baseId}-number`}
           label={schema.title ?? placeholder ?? "Number"}
           hideLabel
-          size="small"
           type="number"
           value={value ?? ""}
           placeholder={placeholder || "Enter number"}
@@ -80,7 +82,6 @@ export function RunAgentInputs({
           id={`${baseId}-textarea`}
           label={schema.title ?? placeholder ?? "Text"}
           hideLabel
-          size="small"
           type="textarea"
           rows={3}
           value={value ?? ""}
@@ -130,7 +131,6 @@ export function RunAgentInputs({
           id={`${baseId}-date`}
           label={schema.title ?? placeholder ?? "Date"}
           hideLabel
-          size="small"
           type="date"
           value={value ? format(value as Date, "yyyy-MM-dd") : ""}
           onChange={(e) => {
@@ -159,7 +159,6 @@ export function RunAgentInputs({
           id={`${baseId}-datetime`}
           label={schema.title ?? placeholder ?? "Date time"}
           hideLabel
-          size="small"
           type="datetime-local"
           value={value ?? ""}
           onChange={(e) => onChange((e.target as HTMLInputElement).value)}
@@ -194,7 +193,6 @@ export function RunAgentInputs({
             label={schema.title ?? placeholder ?? "Select"}
             hideLabel
             value={value ?? ""}
-            size="small"
             onValueChange={(val: string) => onChange(val)}
             placeholder={placeholder || "Select an option"}
             options={schema.enum
@@ -217,7 +215,6 @@ export function RunAgentInputs({
           items={allKeys.map((key) => ({
             value: key,
             label: _schema.properties[key]?.title ?? key,
-            size: "small",
           }))}
           selectedValues={selectedValues}
           onChange={(values: string[]) =>
@@ -336,7 +333,6 @@ export function RunAgentInputs({
           id={`${baseId}-text`}
           label={schema.title ?? placeholder ?? "Text"}
           hideLabel
-          size="small"
           type="text"
           value={value ?? ""}
           onChange={(e) => onChange((e.target as HTMLInputElement).value)}
@@ -347,6 +343,17 @@ export function RunAgentInputs({
   }
 
   return (
-    <div className="no-drag relative flex w-full">{innerInputElement}</div>
+    <div className="flex w-full flex-col gap-0 space-y-2">
+      <label className="large-medium flex items-center gap-1 font-medium">
+        {schema.title || placeholder}
+        <InformationTooltip description={schema.description} />
+      </label>
+      <div
+        className="no-drag relative flex w-full"
+        style={readOnly ? { pointerEvents: "none", opacity: 0.7 } : undefined}
+      >
+        {innerInputElement}
+      </div>
+    </div>
   );
 }

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/modals/RunAgentModal/components/ModalRunSection/ModalRunSection.tsx

@@ -73,22 +73,15 @@ export function ModalRunSection() {
           title="Task Inputs"
           subtitle="Enter the information you want to provide to the agent for this task"
         >
-          {/* Regular inputs */}
           {inputFields.map(([key, inputSubSchema]) => (
-            <div key={key} className="flex w-full flex-col gap-0 space-y-2">
-              <label className="flex items-center gap-1 text-sm font-medium">
-                {inputSubSchema.title || key}
-                <InformationTooltip description={inputSubSchema.description} />
-              </label>
-
-              <RunAgentInputs
-                schema={inputSubSchema}
-                value={inputValues[key] ?? inputSubSchema.default}
-                placeholder={inputSubSchema.description}
-                onChange={(value) => setInputValue(key, value)}
-                data-testid={`agent-input-${key}`}
-              />
-            </div>
+            <RunAgentInputs
+              key={key}
+              schema={inputSubSchema}
+              value={inputValues[key] ?? inputSubSchema.default}
+              placeholder={inputSubSchema.description}
+              onChange={(value) => setInputValue(key, value)}
+              data-testid={`agent-input-${key}`}
+            />
           ))}
         </ModalSection>
       ) : null}

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/selected-views/SelectedRunView/SelectedRunView.tsx

@@ -4,20 +4,19 @@ import { AgentExecutionStatus } from "@/app/api/__generated__/models/agentExecut
 import type { LibraryAgent } from "@/app/api/__generated__/models/libraryAgent";
 import { LoadingSpinner } from "@/components/atoms/LoadingSpinner/LoadingSpinner";
 import { Text } from "@/components/atoms/Text/Text";
-import {
-  Tooltip,
-  TooltipContent,
-  TooltipProvider,
-  TooltipTrigger,
-} from "@/components/atoms/Tooltip/BaseTooltip";
 import { ErrorCard } from "@/components/molecules/ErrorCard/ErrorCard";
+import { InformationTooltip } from "@/components/molecules/InformationTooltip/InformationTooltip";
+import {
+  ScrollableTabs,
+  ScrollableTabsContent,
+  ScrollableTabsList,
+  ScrollableTabsTrigger,
+} from "@/components/molecules/ScrollableTabs/ScrollableTabs";
 import { PendingReviewsList } from "@/components/organisms/PendingReviewsList/PendingReviewsList";
 import { usePendingReviewsForExecution } from "@/hooks/usePendingReviews";
 import { isLargeScreen, useBreakpoint } from "@/lib/hooks/useBreakpoint";
-import { InfoIcon } from "@phosphor-icons/react";
 import { useEffect } from "react";
 import { AgentInputsReadOnly } from "../../modals/AgentInputsReadOnly/AgentInputsReadOnly";
-import { AnchorLinksWrap } from "../AnchorLinksWrap";
 import { LoadingSelectedContent } from "../LoadingSelectedContent";
 import { RunDetailCard } from "../RunDetailCard/RunDetailCard";
 import { RunDetailHeader } from "../RunDetailHeader/RunDetailHeader";
@@ -28,9 +27,6 @@ import { SelectedRunActions } from "./components/SelectedRunActions/SelectedRunA
 import { WebhookTriggerSection } from "./components/WebhookTriggerSection";
 import { useSelectedRunView } from "./useSelectedRunView";
 
-const anchorStyles =
-  "border-b-2 border-transparent pb-1 text-sm font-medium text-slate-600 transition-colors hover:text-slate-900 hover:border-slate-900";
-
 interface Props {
   agent: LibraryAgent;
   runId: string;
@@ -65,13 +61,6 @@ export function SelectedRunView({
   const withSummary = run?.stats?.activity_status;
   const withReviews = run?.status === AgentExecutionStatus.REVIEW;
 
-  function scrollToSection(id: string) {
-    const element = document.getElementById(id);
-    if (element) {
-      element.scrollIntoView({ behavior: "smooth", block: "start" });
-    }
-  }
-
   if (responseError || httpError) {
     return (
       <ErrorCard
@@ -112,118 +101,116 @@ export function SelectedRunView({
                 />
               )}
 
-            {/* Navigation Links */}
-            <AnchorLinksWrap>
-              {withSummary && (
-                <button
-                  onClick={() => scrollToSection("summary")}
-                  className={anchorStyles}
-                >
-                  Summary
-                </button>
-              )}
-              <button
-                onClick={() => scrollToSection("output")}
-                className={anchorStyles}
-              >
-                Output
-              </button>
-              <button
-                onClick={() => scrollToSection("input")}
-                className={anchorStyles}
-              >
-                Your input
-              </button>
-              {withReviews && (
-                <button
-                  onClick={() => scrollToSection("reviews")}
-                  className={anchorStyles}
-                >
-                  Reviews ({pendingReviews.length})
-                </button>
-              )}
-            </AnchorLinksWrap>
-
-            {/* Summary Section */}
-            {withSummary && (
-              <div id="summary" className="scroll-mt-4">
-                <RunDetailCard
-                  title={
-                    <div className="flex items-center gap-2">
-                      <Text variant="lead-semibold">Summary</Text>
-                      <TooltipProvider>
-                        <Tooltip>
-                          <TooltipTrigger asChild>
-                            <InfoIcon
-                              size={16}
-                              className="cursor-help text-neutral-500 hover:text-neutral-700"
-                            />
-                          </TooltipTrigger>
-                          <TooltipContent>
-                            <p className="max-w-xs">
-                              This AI-generated summary describes how the agent
-                              handled your task. It&apos;s an experimental
-                              feature and may occasionally be inaccurate.
-                            </p>
-                          </TooltipContent>
-                        </Tooltip>
-                      </TooltipProvider>
-                    </div>
-                  }
-                >
-                  <RunSummary run={run} />
-                </RunDetailCard>
-              </div>
-            )}
-
-            {/* Output Section */}
-            <div id="output" className="scroll-mt-4">
-              <RunDetailCard title="Output">
-                {isLoading ? (
-                  <div className="text-neutral-500">
-                    <LoadingSpinner />
-                  </div>
-                ) : run && "outputs" in run ? (
-                  <RunOutputs outputs={run.outputs as any} />
-                ) : (
-                  <Text variant="body" className="text-neutral-600">
-                    No output from this run.
-                  </Text>
+            <ScrollableTabs
+              defaultValue="output"
+              className="-mt-2 flex flex-col"
+            >
+              <ScrollableTabsList className="px-4">
+                {withSummary && (
+                  <ScrollableTabsTrigger value="summary">
+                    Summary
+                  </ScrollableTabsTrigger>
                 )}
-              </RunDetailCard>
-            </div>
-
-            {/* Input Section */}
-            <div id="input" className="scroll-mt-4">
-              <RunDetailCard title="Your input">
-                <AgentInputsReadOnly
-                  agent={agent}
-                  inputs={run?.inputs}
-                  credentialInputs={run?.credential_inputs}
-                />
-              </RunDetailCard>
-            </div>
-
-            {/* Reviews Section */}
-            {withReviews && (
-              <div id="reviews" className="scroll-mt-4">
-                <RunDetailCard>
-                  {reviewsLoading ? (
-                    <div className="text-neutral-500">Loading reviews…</div>
-                  ) : pendingReviews.length > 0 ? (
-                    <PendingReviewsList
-                      reviews={pendingReviews}
-                      onReviewComplete={refetchReviews}
-                      emptyMessage="No pending reviews for this execution"
-                    />
-                  ) : (
-                    <div className="text-neutral-600">
-                      No pending reviews for this execution
+                <ScrollableTabsTrigger value="output">
+                  Output
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="input">
+                  Your input
+                </ScrollableTabsTrigger>
+                {withReviews && (
+                  <ScrollableTabsTrigger value="reviews">
+                    Reviews ({pendingReviews.length})
+                  </ScrollableTabsTrigger>
+                )}
+              </ScrollableTabsList>
+              <div className="my-6 flex flex-col gap-6">
+                {/* Summary Section */}
+                {withSummary && (
+                  <ScrollableTabsContent value="summary">
+                    <div className="scroll-mt-4">
+                      <RunDetailCard
+                        title={
+                          <div className="flex items-center gap-1">
+                            <Text variant="lead-semibold">Summary</Text>
+                            <InformationTooltip
+                              iconSize={20}
+                              description="This AI-generated summary describes how the agent handled your task. It's an experimental feature and may occasionally be inaccurate."
+                            />
+                          </div>
+                        }
+                      >
+                        <RunSummary run={run} />
+                      </RunDetailCard>
                     </div>
-                  )}
-                </RunDetailCard>
+                  </ScrollableTabsContent>
+                )}
+
+                {/* Output Section */}
+                <ScrollableTabsContent value="output">
+                  <div className="scroll-mt-4">
+                    <RunDetailCard title="Output">
+                      {isLoading ? (
+                        <div className="text-neutral-500">
+                          <LoadingSpinner />
+                        </div>
+                      ) : run && "outputs" in run ? (
+                        <RunOutputs outputs={run.outputs as any} />
+                      ) : (
+                        <Text variant="body" className="text-neutral-600">
+                          No output from this run.
+                        </Text>
+                      )}
+                    </RunDetailCard>
+                  </div>
+                </ScrollableTabsContent>
+
+                {/* Input Section */}
+                <ScrollableTabsContent value="input">
+                  <div id="input" className="scroll-mt-4">
+                    <RunDetailCard
+                      title={
+                        <div className="flex items-center gap-1">
+                          <Text variant="lead-semibold">Your input</Text>
+                          <InformationTooltip
+                            iconSize={20}
+                            description="This is the input that was provided to the agent for running this task."
+                          />
+                        </div>
+                      }
+                    >
+                      <AgentInputsReadOnly
+                        agent={agent}
+                        inputs={run?.inputs}
+                        credentialInputs={run?.credential_inputs}
+                      />
+                    </RunDetailCard>
+                  </div>
+                </ScrollableTabsContent>
+
+                {/* Reviews Section */}
+                {withReviews && (
+                  <ScrollableTabsContent value="reviews">
+                    <div className="scroll-mt-4">
+                      <RunDetailCard>
+                        {reviewsLoading ? (
+                          <LoadingSpinner size="small" />
+                        ) : pendingReviews.length > 0 ? (
+                          <PendingReviewsList
+                            reviews={pendingReviews}
+                            onReviewComplete={refetchReviews}
+                            emptyMessage="No pending reviews for this execution"
+                          />
+                        ) : (
+                          <Text variant="body" className="text-zinc-700">
+                            No pending reviews for this execution
+                          </Text>
+                        )}
+                      </RunDetailCard>
+                    </div>
+                  </ScrollableTabsContent>
+                )}
               </div>
-            )}
+            </ScrollableTabs>
           </div>
         </SelectedViewLayout>
       </div>

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/selected-views/SelectedScheduleView/SelectedScheduleView.tsx

@@ -9,7 +9,6 @@ import { humanizeCronExpression } from "@/lib/cron-expression-utils";
 import { isLargeScreen, useBreakpoint } from "@/lib/hooks/useBreakpoint";
 import { formatInTimezone, getTimezoneDisplayName } from "@/lib/timezone-utils";
 import { AgentInputsReadOnly } from "../../modals/AgentInputsReadOnly/AgentInputsReadOnly";
-import { AnchorLinksWrap } from "../AnchorLinksWrap";
 import { LoadingSelectedContent } from "../LoadingSelectedContent";
 import { RunDetailCard } from "../RunDetailCard/RunDetailCard";
 import { RunDetailHeader } from "../RunDetailHeader/RunDetailHeader";
@@ -17,9 +16,6 @@ import { SelectedViewLayout } from "../SelectedViewLayout";
 import { SelectedScheduleActions } from "./components/SelectedScheduleActions";
 import { useSelectedScheduleView } from "./useSelectedScheduleView";
 
-const anchorStyles =
-  "border-b-2 border-transparent pb-1 text-sm font-medium text-slate-600 transition-colors hover:text-slate-900 hover:border-slate-900";
-
 interface Props {
   agent: LibraryAgent;
   scheduleId: string;
@@ -45,13 +41,6 @@ export function SelectedScheduleView({
   const breakpoint = useBreakpoint();
   const isLgScreenUp = isLargeScreen(breakpoint);
 
-  function scrollToSection(id: string) {
-    const element = document.getElementById(id);
-    if (element) {
-      element.scrollIntoView({ behavior: "smooth", block: "start" });
-    }
-  }
-
   if (error) {
     return (
       <ErrorCard
@@ -108,22 +97,6 @@ export function SelectedScheduleView({
               ) : null}
             </div>
 
-            {/* Navigation Links */}
-            <AnchorLinksWrap>
-              <button
-                onClick={() => scrollToSection("schedule")}
-                className={anchorStyles}
-              >
-                Schedule
-              </button>
-              <button
-                onClick={() => scrollToSection("input")}
-                className={anchorStyles}
-              >
-                Your input
-              </button>
-            </AnchorLinksWrap>
-
             {/* Schedule Section */}
             <div id="schedule" className="scroll-mt-4">
               <RunDetailCard title="Schedule">

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/selected-views/SelectedScheduleView/components/EditInputsModal/EditInputsModal.tsx

@@ -1,84 +0,0 @@
-"use client";
-
-import type { GraphExecutionJobInfo } from "@/app/api/__generated__/models/graphExecutionJobInfo";
-import type { LibraryAgent } from "@/app/api/__generated__/models/libraryAgent";
-import { Button } from "@/components/atoms/Button/Button";
-import { Text } from "@/components/atoms/Text/Text";
-import { Dialog } from "@/components/molecules/Dialog/Dialog";
-import { PencilSimpleIcon } from "@phosphor-icons/react";
-import { RunAgentInputs } from "../../../../modals/RunAgentInputs/RunAgentInputs";
-import { useEditInputsModal } from "./useEditInputsModal";
-
-type Props = {
-  agent: LibraryAgent;
-  schedule: GraphExecutionJobInfo;
-};
-
-export function EditInputsModal({ agent, schedule }: Props) {
-  const {
-    isOpen,
-    setIsOpen,
-    inputFields,
-    values,
-    setValues,
-    handleSave,
-    isSaving,
-  } = useEditInputsModal(agent, schedule);
-
-  return (
-    <Dialog
-      controlled={{ isOpen, set: setIsOpen }}
-      styling={{ maxWidth: "32rem" }}
-    >
-      <Dialog.Trigger>
-        <Button
-          variant="ghost"
-          size="small"
-          className="absolute -right-2 -top-2"
-        >
-          <PencilSimpleIcon className="size-4" /> Edit inputs
-        </Button>
-      </Dialog.Trigger>
-      <Dialog.Content>
-        <div className="flex flex-col gap-4">
-          <Text variant="h3">Edit inputs</Text>
-          <div className="flex flex-col gap-4">
-            {Object.entries(inputFields).map(([key, fieldSchema]) => (
-              <div key={key} className="flex flex-col gap-1.5">
-                <label className="text-sm font-medium">
-                  {fieldSchema?.title || key}
-                </label>
-                <RunAgentInputs
-                  schema={fieldSchema as any}
-                  value={values[key]}
-                  onChange={(v) => setValues((prev) => ({ ...prev, [key]: v }))}
-                />
-              </div>
-            ))}
-          </div>
-        </div>
-        <Dialog.Footer>
-          <div className="flex w-full justify-end gap-2">
-            <Button
-              variant="secondary"
-              size="small"
-              onClick={() => setIsOpen(false)}
-              className="min-w-32"
-            >
-              Cancel
-            </Button>
-            <Button
-              variant="primary"
-              size="small"
-              onClick={handleSave}
-              loading={isSaving}
-              className="min-w-32"
-            >
-              {isSaving ? "Saving…" : "Save"}
-            </Button>
-          </div>
-        </Dialog.Footer>
-      </Dialog.Content>
-    </Dialog>
-  );
-}

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/selected-views/SelectedScheduleView/components/EditInputsModal/useEditInputsModal.ts

@@ -1,78 +0,0 @@
-"use client";
-
-import { useMemo, useState } from "react";
-import { useQueryClient } from "@tanstack/react-query";
-import { getGetV1ListExecutionSchedulesForAGraphQueryKey } from "@/app/api/__generated__/endpoints/schedules/schedules";
-import type { LibraryAgent } from "@/app/api/__generated__/models/libraryAgent";
-import type { GraphExecutionJobInfo } from "@/app/api/__generated__/models/graphExecutionJobInfo";
-import { useToast } from "@/components/molecules/Toast/use-toast";
-
-function getAgentInputFields(agent: LibraryAgent): Record<string, any> {
-  const schema = agent.input_schema as unknown as {
-    properties?: Record<string, any>;
-  } | null;
-  if (!schema || !schema.properties) return {};
-  const properties = schema.properties as Record<string, any>;
-  const visibleEntries = Object.entries(properties).filter(
-    ([, sub]) => !sub?.hidden,
-  );
-  return Object.fromEntries(visibleEntries);
-}
-
-export function useEditInputsModal(
-  agent: LibraryAgent,
-  schedule: GraphExecutionJobInfo,
-) {
-  const queryClient = useQueryClient();
-  const { toast } = useToast();
-  const [isOpen, setIsOpen] = useState(false);
-  const [isSaving, setIsSaving] = useState(false);
-  const inputFields = useMemo(() => getAgentInputFields(agent), [agent]);
-  const [values, setValues] = useState<Record<string, any>>({
-    ...(schedule.input_data as Record<string, any>),
-  });
-
-  async function handleSave() {
-    setIsSaving(true);
-    try {
-      const res = await fetch(`/api/schedules/${schedule.id}`, {
-        method: "PATCH",
-        headers: { "Content-Type": "application/json" },
-        body: JSON.stringify({ inputs: values }),
-      });
-      if (!res.ok) {
-        let message = "Failed to update schedule inputs";
-        const data = await res.json();
-        message = data?.message || data?.detail || message;
-        throw new Error(message);
-      }
-
-      await queryClient.invalidateQueries({
-        queryKey: getGetV1ListExecutionSchedulesForAGraphQueryKey(
-          schedule.graph_id,
-        ),
-      });
-      toast({
-        title: "Schedule inputs updated",
-      });
-      setIsOpen(false);
-    } catch (error: any) {
-      toast({
-        title: "Failed to update schedule inputs",
-        description: error?.message || "An unexpected error occurred.",
-        variant: "destructive",
-      });
-    }
-    setIsSaving(false);
-  }
-
-  return {
-    isOpen,
-    setIsOpen,
-    inputFields,
-    values,
-    setValues,
-    handleSave,
-    isSaving,
-  } as const;
-}

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/selected-views/SelectedScheduleView/components/SelectedScheduleActions.tsx

@@ -25,9 +25,10 @@ export function SelectedScheduleActions({ agent, scheduleId }: Props) {
           <Button
             variant="icon"
             size="icon"
-            aria-label="Open in builder"
             as="NextLink"
             href={openInBuilderHref}
+            target="_blank"
+            aria-label="View scheduled task details"
           >
             <EyeIcon weight="bold" size={18} className="text-zinc-700" />
           </Button>

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/selected-views/SelectedTemplateView/SelectedTemplateView.tsx

@@ -4,7 +4,6 @@ import type { GraphExecutionMeta } from "@/app/api/__generated__/models/graphExe
 import type { LibraryAgent } from "@/app/api/__generated__/models/libraryAgent";
 import { Input } from "@/components/atoms/Input/Input";
 import { ErrorCard } from "@/components/molecules/ErrorCard/ErrorCard";
-import { InformationTooltip } from "@/components/molecules/InformationTooltip/InformationTooltip";
 import {
   getAgentCredentialsFields,
   getAgentInputFields,
@@ -138,25 +137,13 @@ export function SelectedTemplateView({
               <RunDetailCard title="Your Input">
                 <div className="flex flex-col gap-4">
                   {inputFields.map(([key, inputSubSchema]) => (
-                    <div
+                    <RunAgentInputs
                       key={key}
-                      className="flex w-full flex-col gap-0 space-y-2"
-                    >
-                      <label className="flex items-center gap-1 text-sm font-medium">
-                        {inputSubSchema.title || key}
-                        {inputSubSchema.description && (
-                          <InformationTooltip
-                            description={inputSubSchema.description}
-                          />
-                        )}
-                      </label>
-                      <RunAgentInputs
-                        schema={inputSubSchema}
-                        value={inputs[key] ?? inputSubSchema.default}
-                        placeholder={inputSubSchema.description}
-                        onChange={(value) => setInputValue(key, value)}
-                      />
-                    </div>
+                      schema={inputSubSchema}
+                      value={inputs[key] ?? inputSubSchema.default}
+                      placeholder={inputSubSchema.description}
+                      onChange={(value) => setInputValue(key, value)}
+                    />
                   ))}
                 </div>
               </RunDetailCard>

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/selected-views/SelectedTriggerView/SelectedTriggerView.tsx

@@ -3,7 +3,6 @@
 import type { LibraryAgent } from "@/app/api/__generated__/models/libraryAgent";
 import { Input } from "@/components/atoms/Input/Input";
 import { ErrorCard } from "@/components/molecules/ErrorCard/ErrorCard";
-import { InformationTooltip } from "@/components/molecules/InformationTooltip/InformationTooltip";
 import {
   getAgentCredentialsFields,
   getAgentInputFields,
@@ -131,25 +130,13 @@ export function SelectedTriggerView({
               <RunDetailCard title="Your Input">
                 <div className="flex flex-col gap-4">
                   {inputFields.map(([key, inputSubSchema]) => (
-                    <div
+                    <RunAgentInputs
                       key={key}
-                      className="flex w-full flex-col gap-0 space-y-2"
-                    >
-                      <label className="flex items-center gap-1 text-sm font-medium">
-                        {inputSubSchema.title || key}
-                        {inputSubSchema.description && (
-                          <InformationTooltip
-                            description={inputSubSchema.description}
-                          />
-                        )}
-                      </label>
-                      <RunAgentInputs
-                        schema={inputSubSchema}
-                        value={inputs[key] ?? inputSubSchema.default}
-                        placeholder={inputSubSchema.description}
-                        onChange={(value) => setInputValue(key, value)}
-                      />
-                    </div>
+                      schema={inputSubSchema}
+                      value={inputs[key] ?? inputSubSchema.default}
+                      placeholder={inputSubSchema.description}
+                      onChange={(value) => setInputValue(key, value)}
+                    />
                   ))}
                 </div>
               </RunDetailCard>

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/OldAgentLibraryView/components/agent-run-draft-view.tsx

@@ -680,28 +680,20 @@ export function AgentRunDraftView({
 
             {/* Regular inputs */}
             {Object.entries(agentInputFields).map(([key, inputSubSchema]) => (
-              <div key={key} className="flex flex-col space-y-2">
-                <label className="flex items-center gap-1 text-sm font-medium">
-                  {inputSubSchema.title || key}
-                  <InformationTooltip
-                    description={inputSubSchema.description}
-                  />
-                </label>
-
-                <RunAgentInputs
-                  schema={inputSubSchema}
-                  value={inputValues[key] ?? inputSubSchema.default}
-                  placeholder={inputSubSchema.description}
-                  onChange={(value) => {
-                    setInputValues((obj) => ({
-                      ...obj,
-                      [key]: value,
-                    }));
-                    setChangedPresetAttributes((prev) => prev.add("inputs"));
-                  }}
-                  data-testid={`agent-input-${key}`}
-                />
-              </div>
+              <RunAgentInputs
+                key={key}
+                schema={inputSubSchema}
+                value={inputValues[key] ?? inputSubSchema.default}
+                placeholder={inputSubSchema.description}
+                onChange={(value) => {
+                  setInputValues((obj) => ({
+                    ...obj,
+                    [key]: value,
+                  }));
+                  setChangedPresetAttributes((prev) => prev.add("inputs"));
+                }}
+                data-testid={`agent-input-${key}`}
+              />
             ))}
           </CardContent>
         </Card>

autogpt_platform/frontend/src/app/providers.tsx

@@ -1,36 +1,33 @@
 "use client";
 
-import { LaunchDarklyProvider } from "@/services/feature-flags/feature-flag-provider";
-import OnboardingProvider from "@/providers/onboarding/onboarding-provider";
+import { TooltipProvider } from "@/components/atoms/Tooltip/BaseTooltip";
+import { SentryUserTracker } from "@/components/monitor/SentryUserTracker";
 import { BackendAPIProvider } from "@/lib/autogpt-server-api/context";
 import { getQueryClient } from "@/lib/react-query/queryClient";
-import { QueryClientProvider } from "@tanstack/react-query";
-import {
-  ThemeProvider as NextThemesProvider,
-  ThemeProviderProps,
-} from "next-themes";
-import { NuqsAdapter } from "nuqs/adapters/next/app";
-import { TooltipProvider } from "@/components/atoms/Tooltip/BaseTooltip";
 import CredentialsProvider from "@/providers/agent-credentials/credentials-provider";
-import { SentryUserTracker } from "@/components/monitor/SentryUserTracker";
+import OnboardingProvider from "@/providers/onboarding/onboarding-provider";
+import { LaunchDarklyProvider } from "@/services/feature-flags/feature-flag-provider";
+import { QueryClientProvider } from "@tanstack/react-query";
+import { ThemeProvider, ThemeProviderProps } from "next-themes";
+import { NuqsAdapter } from "nuqs/adapters/next/app";
 
 export function Providers({ children, ...props }: ThemeProviderProps) {
   const queryClient = getQueryClient();
   return (
     <QueryClientProvider client={queryClient}>
       <NuqsAdapter>
-        <NextThemesProvider {...props}>
-          <BackendAPIProvider>
-            <SentryUserTracker />
-            <CredentialsProvider>
-              <LaunchDarklyProvider>
-                <OnboardingProvider>
+        <BackendAPIProvider>
+          <SentryUserTracker />
+          <CredentialsProvider>
+            <LaunchDarklyProvider>
+              <OnboardingProvider>
+                <ThemeProvider forcedTheme="light" {...props}>
                   <TooltipProvider>{children}</TooltipProvider>
-                </OnboardingProvider>
-              </LaunchDarklyProvider>
-            </CredentialsProvider>
-          </BackendAPIProvider>
-        </NextThemesProvider>
+                </ThemeProvider>
+              </OnboardingProvider>
+            </LaunchDarklyProvider>
+          </CredentialsProvider>
+        </BackendAPIProvider>
       </NuqsAdapter>
     </QueryClientProvider>
   );

autogpt_platform/frontend/src/components/molecules/InformationTooltip/InformationTooltip.tsx

@@ -9,16 +9,20 @@ import ReactMarkdown from "react-markdown";
 
 type Props = {
   description?: string;
+  iconSize?: number;
 };
 
-export function InformationTooltip({ description }: Props) {
+export function InformationTooltip({ description, iconSize = 24 }: Props) {
   if (!description) return null;
 
   return (
     <TooltipProvider delayDuration={400}>
       <Tooltip>
         <TooltipTrigger asChild>
-          <Info className="rounded-full p-1 hover:bg-slate-50" size={24} />
+          <Info
+            className="rounded-full p-1 hover:bg-slate-50"
+            size={iconSize}
+          />
         </TooltipTrigger>
         <TooltipContent>
           <ReactMarkdown

autogpt_platform/frontend/src/components/molecules/ScrollableTabs/ScrollableTabs.stories.tsx

@@ -0,0 +1,437 @@
+import type { Meta, StoryObj } from "@storybook/nextjs";
+import {
+  ScrollableTabs,
+  ScrollableTabsContent,
+  ScrollableTabsList,
+  ScrollableTabsTrigger,
+} from "./ScrollableTabs";
+
+const meta = {
+  title: "Molecules/ScrollableTabs",
+  component: ScrollableTabs,
+  parameters: {
+    layout: "fullscreen",
+  },
+  tags: ["autodocs"],
+  argTypes: {},
+} satisfies Meta<typeof ScrollableTabs>;
+
+export default meta;
+type Story = StoryObj<typeof meta>;
+
+function ScrollableTabsDemo() {
+  return (
+    <div className="flex flex-col gap-8 p-8">
+      <h2 className="text-2xl font-bold">ScrollableTabs Examples</h2>
+
+      <div className="space-y-6">
+        <div>
+          <h3 className="mb-4 text-lg font-semibold">
+            Short Content (Tabs Hidden)
+          </h3>
+          <div className="h-[300px] overflow-y-auto border border-zinc-200">
+            <ScrollableTabs defaultValue="tab1" className="h-full">
+              <ScrollableTabsList>
+                <ScrollableTabsTrigger value="tab1">
+                  Account
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="tab2">
+                  Password
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="tab3">
+                  Settings
+                </ScrollableTabsTrigger>
+              </ScrollableTabsList>
+              <ScrollableTabsContent value="tab1">
+                <div className="p-4 text-sm">
+                  Make changes to your account here. Click save when you&apos;re
+                  done.
+                </div>
+              </ScrollableTabsContent>
+              <ScrollableTabsContent value="tab2">
+                <div className="p-4 text-sm">
+                  Change your password here. After saving, you&apos;ll be logged
+                  out.
+                </div>
+              </ScrollableTabsContent>
+              <ScrollableTabsContent value="tab3">
+                <div className="p-4 text-sm">
+                  Update your preferences and settings here.
+                </div>
+              </ScrollableTabsContent>
+            </ScrollableTabs>
+          </div>
+        </div>
+
+        <div>
+          <h3 className="mb-4 text-lg font-semibold">
+            Long Content (Tabs Visible)
+          </h3>
+          <div className="h-[400px] overflow-y-auto border border-zinc-200">
+            <ScrollableTabs defaultValue="tab1" className="h-full">
+              <ScrollableTabsList>
+                <ScrollableTabsTrigger value="tab1">
+                  Account
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="tab2">
+                  Password
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="tab3">
+                  Settings
+                </ScrollableTabsTrigger>
+              </ScrollableTabsList>
+              <ScrollableTabsContent value="tab1">
+                <div className="p-8 text-sm">
+                  <h4 className="mb-4 text-lg font-semibold">
+                    Account Settings
+                  </h4>
+                  <p className="mb-4">
+                    Make changes to your account here. Click save when
+                    you&apos;re done.
+                  </p>
+                  <p className="mb-4">
+                    Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed
+                    do eiusmod tempor incididunt ut labore et dolore magna
+                    aliqua. Ut enim ad minim veniam, quis nostrud exercitation
+                    ullamco laboris.
+                  </p>
+                  <p className="mb-4">
+                    Duis aute irure dolor in reprehenderit in voluptate velit
+                    esse cillum dolore eu fugiat nulla pariatur. Excepteur sint
+                    occaecat cupidatat non proident.
+                  </p>
+                  <p>
+                    Sed ut perspiciatis unde omnis iste natus error sit
+                    voluptatem accusantium doloremque laudantium, totam rem
+                    aperiam.
+                  </p>
+                </div>
+              </ScrollableTabsContent>
+              <ScrollableTabsContent value="tab2">
+                <div className="p-8 text-sm">
+                  <h4 className="mb-4 text-lg font-semibold">
+                    Password Settings
+                  </h4>
+                  <p className="mb-4">
+                    Change your password here. After saving, you&apos;ll be
+                    logged out.
+                  </p>
+                  <p className="mb-4">
+                    At vero eos et accusamus et iusto odio dignissimos ducimus
+                    qui blanditiis praesentium voluptatum deleniti atque
+                    corrupti quos dolores et quas molestias excepturi sint
+                    occaecati cupiditate.
+                  </p>
+                  <p className="mb-4">
+                    Et harum quidem rerum facilis est et expedita distinctio.
+                    Nam libero tempore, cum soluta nobis est eligendi optio
+                    cumque nihil impedit quo minus.
+                  </p>
+                  <p>
+                    Temporibus autem quibusdam et aut officiis debitis aut rerum
+                    necessitatibus saepe eveniet ut et voluptates repudiandae
+                    sint.
+                  </p>
+                </div>
+              </ScrollableTabsContent>
+              <ScrollableTabsContent value="tab3">
+                <div className="p-8 text-sm">
+                  <h4 className="mb-4 text-lg font-semibold">
+                    General Settings
+                  </h4>
+                  <p className="mb-4">
+                    Update your preferences and settings here.
+                  </p>
+                  <p className="mb-4">
+                    Nemo enim ipsam voluptatem quia voluptas sit aspernatur aut
+                    odit aut fugit, sed quia consequuntur magni dolores eos qui
+                    ratione voluptatem sequi nesciunt.
+                  </p>
+                  <p className="mb-4">
+                    Neque porro quisquam est, qui dolorem ipsum quia dolor sit
+                    amet, consectetur, adipisci velit, sed quia non numquam eius
+                    modi tempora incidunt ut labore et dolore magnam aliquam
+                    quaerat voluptatem.
+                  </p>
+                  <p>
+                    Ut enim ad minima veniam, quis nostrum exercitationem ullam
+                    corporis suscipit laboriosam, nisi ut aliquid ex ea commodi
+                    consequatur.
+                  </p>
+                </div>
+              </ScrollableTabsContent>
+            </ScrollableTabs>
+          </div>
+        </div>
+
+        <div>
+          <h3 className="mb-4 text-lg font-semibold">Many Tabs</h3>
+          <div className="h-[500px] overflow-y-auto border border-zinc-200">
+            <ScrollableTabs defaultValue="overview" className="h-full">
+              <ScrollableTabsList>
+                <ScrollableTabsTrigger value="overview">
+                  Overview
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="analytics">
+                  Analytics
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="reports">
+                  Reports
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="notifications">
+                  Notifications
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="integrations">
+                  Integrations
+                </ScrollableTabsTrigger>
+                <ScrollableTabsTrigger value="billing">
+                  Billing
+                </ScrollableTabsTrigger>
+              </ScrollableTabsList>
+              <ScrollableTabsContent value="overview">
+                <div className="p-8 text-sm">
+                  <h4 className="mb-4 text-lg font-semibold">
+                    Dashboard Overview
+                  </h4>
+                  <p className="mb-4">
+                    Dashboard overview with key metrics and recent activity.
+                  </p>
+                  <p className="mb-4">
+                    Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed
+                    do eiusmod tempor incididunt ut labore et dolore magna
+                    aliqua.
+                  </p>
+                  <p>
+                    Ut enim ad minim veniam, quis nostrud exercitation ullamco
+                    laboris nisi ut aliquip ex ea commodo consequat.
+                  </p>
+                </div>
+              </ScrollableTabsContent>
+              <ScrollableTabsContent value="analytics">
+                <div className="p-8 text-sm">
+                  <h4 className="mb-4 text-lg font-semibold">Analytics</h4>
+                  <p className="mb-4">
+                    Detailed analytics and performance metrics.
+                  </p>
+                  <p className="mb-4">
+                    Duis aute irure dolor in reprehenderit in voluptate velit
+                    esse cillum dolore eu fugiat nulla pariatur.
+                  </p>
+                  <p>
+                    Excepteur sint occaecat cupidatat non proident, sunt in
+                    culpa qui officia deserunt mollit anim id est laborum.
+                  </p>
+                </div>
+              </ScrollableTabsContent>
+              <ScrollableTabsContent value="reports">
+                <div className="p-8 text-sm">
+                  <h4 className="mb-4 text-lg font-semibold">Reports</h4>
+                  <p className="mb-4">
+                    Generate and view reports for your account.
+                  </p>
+                  <p className="mb-4">
+                    Sed ut perspiciatis unde omnis iste natus error sit
+                    voluptatem accusantium doloremque laudantium.
+                  </p>
+                  <p>
+                    Totam rem aperiam, eaque ipsa quae ab illo inventore
+                    veritatis et quasi architecto beatae vitae dicta sunt
+                    explicabo.
+                  </p>
+                </div>
+              </ScrollableTabsContent>
+              <ScrollableTabsContent value="notifications">
+                <div className="p-8 text-sm">
+                  <h4 className="mb-4 text-lg font-semibold">Notifications</h4>
+                  <p className="mb-4">Manage your notification preferences.</p>
+                  <p className="mb-4">
+                    Nemo enim ipsam voluptatem quia voluptas sit aspernatur aut
+                    odit aut fugit.
+                  </p>
+                  <p>
+                    Sed quia consequuntur magni dolores eos qui ratione
+                    voluptatem sequi nesciunt.
+                  </p>
+                </div>
+              </ScrollableTabsContent>
+              <ScrollableTabsContent value="integrations">
+                <div className="p-8 text-sm">
+                  <h4 className="mb-4 text-lg font-semibold">Integrations</h4>
+                  <p className="mb-4">
+                    Connect and manage third-party integrations.
+                  </p>
+                  <p className="mb-4">
+                    Neque porro quisquam est, qui dolorem ipsum quia dolor sit
+                    amet.
+                  </p>
+                  <p>
+                    Consectetur, adipisci velit, sed quia non numquam eius modi
+                    tempora incidunt.
+                  </p>
+                </div>
+              </ScrollableTabsContent>
+              <ScrollableTabsContent value="billing">
+                <div className="p-8 text-sm">
+                  <h4 className="mb-4 text-lg font-semibold">Billing</h4>
+                  <p className="mb-4">
+                    View and manage your billing information.
+                  </p>
+                  <p className="mb-4">
+                    Ut enim ad minima veniam, quis nostrum exercitationem ullam
+                    corporis suscipit laboriosam.
+                  </p>
+                  <p>
+                    Nisi ut aliquid ex ea commodi consequatur? Quis autem vel
+                    eum iure reprehenderit qui in ea voluptate velit esse.
+                  </p>
+                </div>
+              </ScrollableTabsContent>
+            </ScrollableTabs>
+          </div>
+        </div>
+      </div>
+    </div>
+  );
+}
+
+export const Default = {
+  render: () => <ScrollableTabsDemo />,
+} satisfies Story;
+
+export const ShortContent = {
+  render: () => (
+    <div className="p-8">
+      <div className="h-[200px] overflow-y-auto border border-zinc-200">
+        <ScrollableTabs defaultValue="account" className="h-full">
+          <ScrollableTabsList>
+            <ScrollableTabsTrigger value="account">
+              Account
+            </ScrollableTabsTrigger>
+            <ScrollableTabsTrigger value="password">
+              Password
+            </ScrollableTabsTrigger>
+          </ScrollableTabsList>
+          <ScrollableTabsContent value="account">
+            <div className="p-4 text-sm">
+              Make changes to your account here. Click save when you&apos;re
+              done.
+            </div>
+          </ScrollableTabsContent>
+          <ScrollableTabsContent value="password">
+            <div className="p-4 text-sm">
+              Change your password here. After saving, you&apos;ll be logged
+              out.
+            </div>
+          </ScrollableTabsContent>
+        </ScrollableTabs>
+      </div>
+    </div>
+  ),
+} satisfies Story;
+
+export const LongContent = {
+  render: () => (
+    <div className="p-8">
+      <div className="h-[600px] overflow-y-auto border border-zinc-200">
+        <ScrollableTabs defaultValue="tab1" className="h-full">
+          <ScrollableTabsList>
+            <ScrollableTabsTrigger value="tab1">Account</ScrollableTabsTrigger>
+            <ScrollableTabsTrigger value="tab2">Password</ScrollableTabsTrigger>
+            <ScrollableTabsTrigger value="tab3">Settings</ScrollableTabsTrigger>
+          </ScrollableTabsList>
+          <ScrollableTabsContent value="tab1">
+            <div className="p-8 text-sm">
+              <h4 className="mb-4 text-lg font-semibold">Account Settings</h4>
+              <p className="mb-4">
+                Make changes to your account here. Click save when you&apos;re
+                done.
+              </p>
+              <p className="mb-4">
+                Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do
+                eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut
+                enim ad minim veniam, quis nostrud exercitation ullamco laboris
+                nisi ut aliquip ex ea commodo consequat.
+              </p>
+              <p className="mb-4">
+                Duis aute irure dolor in reprehenderit in voluptate velit esse
+                cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat
+                cupidatat non proident, sunt in culpa qui officia deserunt
+                mollit anim id est laborum.
+              </p>
+              <p className="mb-4">
+                Sed ut perspiciatis unde omnis iste natus error sit voluptatem
+                accusantium doloremque laudantium, totam rem aperiam, eaque ipsa
+                quae ab illo inventore veritatis et quasi architecto beatae
+                vitae dicta sunt explicabo.
+              </p>
+              <p>
+                Nemo enim ipsam voluptatem quia voluptas sit aspernatur aut odit
+                aut fugit, sed quia consequuntur magni dolores eos qui ratione
+                voluptatem sequi nesciunt.
+              </p>
+            </div>
+          </ScrollableTabsContent>
+          <ScrollableTabsContent value="tab2">
+            <div className="p-8 text-sm">
+              <h4 className="mb-4 text-lg font-semibold">Password Settings</h4>
+              <p className="mb-4">
+                Change your password here. After saving, you&apos;ll be logged
+                out.
+              </p>
+              <p className="mb-4">
+                At vero eos et accusamus et iusto odio dignissimos ducimus qui
+                blanditiis praesentium voluptatum deleniti atque corrupti quos
+                dolores et quas molestias excepturi sint occaecati cupiditate
+                non provident.
+              </p>
+              <p className="mb-4">
+                Similique sunt in culpa qui officia deserunt mollitia animi, id
+                est laborum et dolorum fuga. Et harum quidem rerum facilis est
+                et expedita distinctio.
+              </p>
+              <p className="mb-4">
+                Nam libero tempore, cum soluta nobis est eligendi optio cumque
+                nihil impedit quo minus id quod maxime placeat facere possimus,
+                omnis voluptas assumenda est, omnis dolor repellendus.
+              </p>
+              <p>
+                Temporibus autem quibusdam et aut officiis debitis aut rerum
+                necessitatibus saepe eveniet ut et voluptates repudiandae sint
+                et molestiae non recusandae.
+              </p>
+            </div>
+          </ScrollableTabsContent>
+          <ScrollableTabsContent value="tab3">
+            <div className="p-8 text-sm">
+              <h4 className="mb-4 text-lg font-semibold">General Settings</h4>
+              <p className="mb-4">Update your preferences and settings here.</p>
+              <p className="mb-4">
+                Neque porro quisquam est, qui dolorem ipsum quia dolor sit amet,
+                consectetur, adipisci velit, sed quia non numquam eius modi
+                tempora incidunt ut labore et dolore magnam aliquam quaerat
+                voluptatem.
+              </p>
+              <p className="mb-4">
+                Ut enim ad minima veniam, quis nostrum exercitationem ullam
+                corporis suscipit laboriosam, nisi ut aliquid ex ea commodi
+                consequatur? Quis autem vel eum iure reprehenderit qui in ea
+                voluptate velit esse quam nihil molestiae consequatur.
+              </p>
+              <p className="mb-4">
+                Vel illum qui dolorem eum fugiat quo voluptas nulla pariatur? At
+                vero eos et accusamus et iusto odio dignissimos ducimus qui
+                blanditiis praesentium voluptatum deleniti atque corrupti quos
+                dolores.
+              </p>
+              <p>
+                Et quas molestias excepturi sint occaecati cupiditate non
+                provident, similique sunt in culpa qui officia deserunt mollitia
+                animi, id est laborum et dolorum fuga.
+              </p>
+            </div>
+          </ScrollableTabsContent>
+        </ScrollableTabs>
+      </div>
+    </div>
+  ),
+} satisfies Story;

autogpt_platform/frontend/src/components/molecules/ScrollableTabs/ScrollableTabs.tsx

@@ -0,0 +1,59 @@
+"use client";
+
+import { cn } from "@/lib/utils";
+import { Children } from "react";
+import { ScrollableTabsContent } from "./components/ScrollableTabsContent";
+import { ScrollableTabsList } from "./components/ScrollableTabsList";
+import { ScrollableTabsTrigger } from "./components/ScrollableTabsTrigger";
+import { ScrollableTabsContext } from "./context";
+import { findContentElements, findListElement } from "./helpers";
+import { useScrollableTabsInternal } from "./useScrollableTabs";
+
+interface Props {
+  children?: React.ReactNode;
+  className?: string;
+  defaultValue?: string;
+}
+
+export function ScrollableTabs({ children, className, defaultValue }: Props) {
+  const {
+    activeValue,
+    setActiveValue,
+    registerContent,
+    scrollToSection,
+    scrollContainer,
+    contentContainerRef,
+  } = useScrollableTabsInternal({ defaultValue });
+
+  const childrenArray = Children.toArray(children);
+  const listElement = findListElement(childrenArray);
+  const contentElements = findContentElements(childrenArray);
+
+  return (
+    <ScrollableTabsContext.Provider
+      value={{
+        activeValue,
+        setActiveValue,
+        registerContent,
+        scrollToSection,
+        scrollContainer,
+      }}
+    >
+      <div className={cn("relative flex flex-col", className)}>
+        {listElement}
+        <div
+          ref={(node) => {
+            if (contentContainerRef) {
+              contentContainerRef.current = node;
+            }
+          }}
+          className="max-h-[64rem] overflow-y-auto scrollbar-thin scrollbar-track-transparent scrollbar-thumb-zinc-300 dark:scrollbar-thumb-zinc-700"
+        >
+          <div className="min-h-full pb-[200px]">{contentElements}</div>
+        </div>
+      </div>
+    </ScrollableTabsContext.Provider>
+  );
+}
+
+export { ScrollableTabsContent, ScrollableTabsList, ScrollableTabsTrigger };

autogpt_platform/frontend/src/components/molecules/ScrollableTabs/components/ScrollableTabsContent.tsx

@@ -0,0 +1,48 @@
+"use client";
+
+import { cn } from "@/lib/utils";
+import * as React from "react";
+import { useScrollableTabs } from "../context";
+
+interface Props extends React.HTMLAttributes<HTMLDivElement> {
+  value: string;
+}
+
+export const ScrollableTabsContent = React.forwardRef<HTMLDivElement, Props>(
+  function ScrollableTabsContent(
+    { className, value, children, ...props },
+    ref,
+  ) {
+    const { registerContent } = useScrollableTabs();
+    const contentRef = React.useRef<HTMLDivElement>(null);
+
+    React.useEffect(() => {
+      if (contentRef.current) {
+        registerContent(value, contentRef.current);
+      }
+      return () => {
+        registerContent(value, null);
+      };
+    }, [value, registerContent]);
+
+    return (
+      <div
+        ref={(node) => {
+          if (typeof ref === "function") ref(node);
+          else if (ref) ref.current = node;
+          // eslint-disable-next-line @typescript-eslint/ban-ts-comment
+          // @ts-ignore
+          contentRef.current = node;
+        }}
+        data-scrollable-tab-content
+        data-value={value}
+        className={cn("focus-visible:outline-none", className)}
+        {...props}
+      >
+        {children}
+      </div>
+    );
+  },
+);
+
+ScrollableTabsContent.displayName = "ScrollableTabsContent";

autogpt_platform/frontend/src/components/molecules/ScrollableTabs/components/ScrollableTabsList.tsx

@@ -0,0 +1,52 @@
+"use client";
+
+import { cn } from "@/lib/utils";
+import * as React from "react";
+import { useScrollableTabs } from "../context";
+
+export const ScrollableTabsList = React.forwardRef<
+  HTMLDivElement,
+  React.HTMLAttributes<HTMLDivElement>
+>(function ScrollableTabsList({ className, children, ...props }, ref) {
+  const { activeValue } = useScrollableTabs();
+  const [activeTabElement, setActiveTabElement] =
+    React.useState<HTMLElement | null>(null);
+
+  React.useEffect(() => {
+    const activeButton = Array.from(
+      document.querySelectorAll<HTMLElement>(
+        '[data-scrollable-tab-trigger][data-value="' + activeValue + '"]',
+      ),
+    )[0];
+
+    if (activeButton) {
+      setActiveTabElement(activeButton);
+    }
+  }, [activeValue]);
+
+  return (
+    <div className="relative" ref={ref}>
+      <div
+        className={cn(
+          "inline-flex w-full items-center justify-start border-b border-zinc-100",
+          className,
+        )}
+        {...props}
+      >
+        {children}
+      </div>
+      {activeTabElement && (
+        <div
+          className="transition-left transition-right absolute bottom-0 h-0.5 bg-purple-600 duration-200 ease-in-out"
+          style={{
+            left: activeTabElement.offsetLeft,
+            width: activeTabElement.offsetWidth,
+            willChange: "left, width",
+          }}
+        />
+      )}
+    </div>
+  );
+});
+
+ScrollableTabsList.displayName = "ScrollableTabsList";

autogpt_platform/frontend/src/components/molecules/ScrollableTabs/components/ScrollableTabsTrigger.tsx

@@ -0,0 +1,53 @@
+"use client";
+
+import { cn } from "@/lib/utils";
+import * as React from "react";
+import { useScrollableTabs } from "../context";
+
+interface Props extends React.ButtonHTMLAttributes<HTMLButtonElement> {
+  value: string;
+}
+
+export const ScrollableTabsTrigger = React.forwardRef<HTMLButtonElement, Props>(
+  function ScrollableTabsTrigger(
+    { className, value, children, ...props },
+    ref,
+  ) {
+    const { activeValue, scrollToSection } = useScrollableTabs();
+    const elementRef = React.useRef<HTMLButtonElement>(null);
+    const isActive = activeValue === value;
+
+    function handleClick(e: React.MouseEvent<HTMLButtonElement>) {
+      e.preventDefault();
+      e.stopPropagation();
+      scrollToSection(value);
+      props.onClick?.(e);
+    }
+
+    return (
+      <button
+        type="button"
+        ref={(node) => {
+          if (typeof ref === "function") ref(node);
+          else if (ref) ref.current = node;
+          // eslint-disable-next-line @typescript-eslint/ban-ts-comment
+          // @ts-ignore
+          elementRef.current = node;
+        }}
+        data-scrollable-tab-trigger
+        data-value={value}
+        onClick={handleClick}
+        className={cn(
+          "relative inline-flex items-center justify-center whitespace-nowrap px-3 py-3 font-sans text-[0.875rem] font-medium leading-[1.5rem] text-zinc-700 transition-all focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-neutral-400 focus-visible:ring-offset-2 disabled:pointer-events-none disabled:opacity-50",
+          isActive && "text-purple-600",
+          className,
+        )}
+        {...props}
+      >
+        {children}
+      </button>
+    );
+  },
+);
+
+ScrollableTabsTrigger.displayName = "ScrollableTabsTrigger";

autogpt_platform/frontend/src/components/molecules/ScrollableTabs/context.ts

@@ -0,0 +1,22 @@
+import * as React from "react";
+import { createContext, useContext } from "react";
+
+interface ScrollableTabsContextValue {
+  activeValue: string | null;
+  setActiveValue: React.Dispatch<React.SetStateAction<string | null>>;
+  registerContent: (value: string, element: HTMLElement | null) => void;
+  scrollToSection: (value: string) => void;
+  scrollContainer: HTMLElement | null;
+}
+
+export const ScrollableTabsContext = createContext<
+  ScrollableTabsContextValue | undefined
+>(undefined);
+
+export function useScrollableTabs() {
+  const context = useContext(ScrollableTabsContext);
+  if (!context) {
+    throw new Error("useScrollableTabs must be used within a ScrollableTabs");
+  }
+  return context;
+}

autogpt_platform/frontend/src/components/molecules/ScrollableTabs/helpers.ts

@@ -0,0 +1,48 @@
+import * as React from "react";
+
+const HEADER_OFFSET = 100;
+
+export function calculateScrollPosition(
+  elementRect: DOMRect,
+  containerRect: DOMRect,
+  currentScrollTop: number,
+): number {
+  const elementTopRelativeToContainer =
+    elementRect.top - containerRect.top + currentScrollTop - HEADER_OFFSET;
+
+  return Math.max(0, elementTopRelativeToContainer);
+}
+
+function hasDisplayName(
+  type: unknown,
+  displayName: string,
+): type is { displayName: string } {
+  return (
+    typeof type === "object" &&
+    type !== null &&
+    "displayName" in type &&
+    (type as { displayName: unknown }).displayName === displayName
+  );
+}
+
+export function findListElement(
+  children: React.ReactNode[],
+): React.ReactElement | undefined {
+  return children.find(
+    (child) =>
+      React.isValidElement(child) &&
+      hasDisplayName(child.type, "ScrollableTabsList"),
+  ) as React.ReactElement | undefined;
+}
+
+export function findContentElements(
+  children: React.ReactNode[],
+): React.ReactNode[] {
+  return children.filter(
+    (child) =>
+      !(
+        React.isValidElement(child) &&
+        hasDisplayName(child.type, "ScrollableTabsList")
+      ),
+  );
+}

autogpt_platform/frontend/src/components/molecules/ScrollableTabs/useScrollableTabs.ts

@@ -0,0 +1,60 @@
+import { useCallback, useRef, useState } from "react";
+import { calculateScrollPosition } from "./helpers";
+
+interface Args {
+  defaultValue?: string;
+}
+
+export function useScrollableTabsInternal({ defaultValue }: Args) {
+  const [activeValue, setActiveValue] = useState<string | null>(
+    defaultValue || null,
+  );
+  const contentRefs = useRef<Map<string, HTMLElement>>(new Map());
+  const contentContainerRef = useRef<HTMLDivElement | null>(null);
+
+  function registerContent(value: string, element: HTMLElement | null) {
+    if (element) {
+      contentRefs.current.set(value, element);
+    } else {
+      contentRefs.current.delete(value);
+    }
+  }
+
+  function scrollToSection(value: string) {
+    const element = contentRefs.current.get(value);
+    const scrollContainer = contentContainerRef.current;
+    if (!element || !scrollContainer) return;
+
+    setActiveValue(value);
+
+    const containerRect = scrollContainer.getBoundingClientRect();
+    const elementRect = element.getBoundingClientRect();
+    const currentScrollTop = scrollContainer.scrollTop;
+    const scrollTop = calculateScrollPosition(
+      elementRect,
+      containerRect,
+      currentScrollTop,
+    );
+
+    const maxScrollTop =
+      scrollContainer.scrollHeight - scrollContainer.clientHeight;
+    const clampedScrollTop = Math.min(Math.max(0, scrollTop), maxScrollTop);
+
+    scrollContainer.scrollTo({
+      top: clampedScrollTop,
+      behavior: "smooth",
+    });
+  }
+
+  const memoizedRegisterContent = useCallback(registerContent, []);
+  const memoizedScrollToSection = useCallback(scrollToSection, []);
+
+  return {
+    activeValue,
+    setActiveValue,
+    registerContent: memoizedRegisterContent,
+    scrollToSection: memoizedScrollToSection,
+    scrollContainer: contentContainerRef.current,
+    contentContainerRef,
+  };
+}

autogpt_platform/frontend/tailwind.config.ts

@@ -1,10 +1,10 @@
+import scrollbar from "tailwind-scrollbar";
 import type { Config } from "tailwindcss";
 import tailwindcssAnimate from "tailwindcss-animate";
-import scrollbar from "tailwind-scrollbar";
 import { colors } from "./src/components/styles/colors";
 
 const config = {
-  darkMode: ["class"],
+  darkMode: ["class", ".dark-mode"], // ignore dark: prefix classes for now until we fully support dark mode
   content: ["./src/**/*.{ts,tsx}"],
   prefix: "",
   theme: {