fix(classic): always use native tool calling, fix N/A command loop

- Remove openai_functions config option - native tool calling is now always enabled
- Remove use_functions_api from BaseAgentConfiguration and prompt strategy
- Add use_prefill config to disable prefill for Anthropic (prefill + tools incompatible)
- Update anthropic dependency to ^0.45.0 for tools API support
- Simplify prompt strategy to always expect tool_calls from LLM response

This fixes the N/A command loop bug where models would output "N/A" as a
command name when function calling was disabled. With native tool calling
always enabled, models are forced to pick from valid tools only.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

classic/forge/CLAUDE.md

@@ -102,7 +102,6 @@ smart_llm: ModelName = "gpt-4"
 big_brain: bool = True              # Use smart_llm
 cycle_budget: Optional[int] = 1     # Steps before approval needed
 send_token_limit: Optional[int]     # Prompt token budget
-use_functions_api: bool = False
 ```
 
 ### Component System (`agent/components.py`)

classic/forge/forge/agent/base.py

@@ -18,7 +18,7 @@ from typing import (
 )
 
 from colorama import Fore
-from pydantic import BaseModel, Field, ValidationInfo, field_validator
+from pydantic import BaseModel, Field
 from pydantic_core import from_json, to_json
 
 from forge.agent import protocols
@@ -53,7 +53,6 @@ class BaseAgentConfiguration(SystemConfiguration):
 
     fast_llm: ModelName = UserConfigurable(default=OpenAIModelName.GPT3_16k)
     smart_llm: ModelName = UserConfigurable(default=OpenAIModelName.GPT4)
-    use_functions_api: bool = UserConfigurable(default=False)
 
     default_cycle_instruction: str = DEFAULT_TRIGGERING_PROMPT
     """The default instruction passed to the AI for a thinking cycle."""
@@ -85,22 +84,6 @@ class BaseAgentConfiguration(SystemConfiguration):
     defaults to 75% of `llm.max_tokens`.
     """
 
-    @field_validator("use_functions_api")
-    def validate_openai_functions(cls, value: bool, info: ValidationInfo):
-        if value:
-            smart_llm = info.data["smart_llm"]
-            fast_llm = info.data["fast_llm"]
-            assert all(
-                [
-                    not any(s in name for s in {"-0301", "-0314"})
-                    for name in {smart_llm, fast_llm}
-                ]
-            ), (
-                f"Model {smart_llm} does not support OpenAI Functions. "
-                "Please disable OPENAI_FUNCTIONS or choose a suitable model."
-            )
-        return value
-
 
 class BaseAgentSettings(SystemSettings):
     agent_id: str = ""

classic/forge/pyproject.toml

@@ -27,7 +27,7 @@ python = "^3.10"
 agbenchmark = { path = "../benchmark", optional = true }
 # agbenchmark = {git = "https://github.com/Significant-Gravitas/AutoGPT.git", subdirectory = "benchmark", optional = true}
 aiohttp = "^3.8.5"
-anthropic = ">=0.40,<1.0"
+anthropic = "^0.45.0"
 beautifulsoup4 = "^4.12.2"
 boto3 = "^1.33.6"
 charset-normalizer = "^3.1.0"

classic/original_autogpt/autogpt/agent_factory/configurators.py

@@ -109,7 +109,6 @@ def create_agent_state(
             fast_llm=app_config.fast_llm,
             smart_llm=app_config.smart_llm,
             allow_fs_access=not app_config.restrict_to_workspace,
-            use_functions_api=app_config.openai_functions,
         ),
         history=Agent.default_settings.history.model_copy(deep=True),
     )

classic/original_autogpt/autogpt/agents/agent.py

@@ -5,8 +5,6 @@ import logging
 from typing import TYPE_CHECKING, Any, ClassVar, Optional
 
 import sentry_sdk
-from pydantic import Field
-
 from forge.agent.base import BaseAgent, BaseAgentConfiguration, BaseAgentSettings
 from forge.agent.protocols import (
     AfterExecute,
@@ -64,6 +62,7 @@ from forge.utils.exceptions import (
     CommandExecutionError,
     UnknownCommandError,
 )
+from pydantic import Field
 
 from .prompt_strategies.one_shot import (
     OneShotAgentActionProposal,
@@ -113,11 +112,8 @@ class Agent(BaseAgent[OneShotAgentActionProposal], Configurable[AgentSettings]):
         prompt_config = OneShotAgentPromptStrategy.default_configuration.model_copy(
             deep=True
         )
-        prompt_config.use_functions_api = (
-            settings.config.use_functions_api
-            # Anthropic currently doesn't support tools + prefilling :(
-            and self.llm.provider_name != "anthropic"
-        )
+        # Anthropic doesn't support tools + prefilling, so disable prefill for Anthropic
+        prompt_config.use_prefill = self.llm.provider_name != "anthropic"
         self.prompt_strategy = OneShotAgentPromptStrategy(prompt_config, logger)
         self.commands: list[Command] = []
 

classic/original_autogpt/autogpt/agents/prompt_strategies/one_shot.py

@@ -73,7 +73,7 @@ class OneShotAgentPromptConfiguration(SystemConfiguration):
     choose_action_instruction: str = UserConfigurable(
         default=DEFAULT_CHOOSE_ACTION_INSTRUCTION
     )
-    use_functions_api: bool = UserConfigurable(default=False)
+    use_prefill: bool = True
 
     #########
     # State #
@@ -134,8 +134,8 @@ class OneShotAgentPromptStrategy(PromptStrategy):
                 *messages,
                 final_instruction_msg,
             ],
-            prefill_response=response_prefill,
-            functions=commands if self.config.use_functions_api else [],
+            prefill_response=response_prefill if self.config.use_prefill else "",
+            functions=commands,
         )
 
     def build_system_prompt(
@@ -152,9 +152,7 @@ class OneShotAgentPromptStrategy(PromptStrategy):
             str: The system prompt body
             str: The desired start for the LLM's response; used to steer the output
         """
-        response_fmt_instruction, response_prefill = self.response_format_instruction(
-            self.config.use_functions_api
-        )
+        response_fmt_instruction, response_prefill = self.response_format_instruction()
         system_prompt_parts = (
             self._generate_intro_prompt(ai_profile)
             + (self._generate_os_info() if include_os_info else [])
@@ -181,10 +179,11 @@ class OneShotAgentPromptStrategy(PromptStrategy):
             response_prefill,
         )
 
-    def response_format_instruction(self, use_functions_api: bool) -> tuple[str, str]:
+    def response_format_instruction(self) -> tuple[str, str]:
         response_schema = self.response_schema.model_copy(deep=True)
         assert response_schema.properties
-        if use_functions_api and "use_tool" in response_schema.properties:
+        # Always use tool calling - remove use_tool from schema since it comes from tool_calls
+        if "use_tool" in response_schema.properties:
             del response_schema.properties["use_tool"]
 
         # Unindent for performance
@@ -199,7 +198,7 @@ class OneShotAgentPromptStrategy(PromptStrategy):
             (
                 f"YOU MUST ALWAYS RESPOND WITH A JSON OBJECT OF THE FOLLOWING TYPE:\n"
                 f"{response_format}"
-                + ("\n\nYOU MUST ALSO INVOKE A TOOL!" if use_functions_api else "")
+                "\n\nYOU MUST ALSO INVOKE A TOOL!"
             ),
             response_prefill,
         )
@@ -269,13 +268,13 @@ class OneShotAgentPromptStrategy(PromptStrategy):
             "Parsing object extracted from LLM response:\n"
             f"{json.dumps(assistant_reply_dict, indent=4)}"
         )
-        if self.config.use_functions_api:
-            if not response.tool_calls:
-                raise InvalidAgentResponseError("Assistant did not use a tool")
-            assistant_reply_dict["use_tool"] = response.tool_calls[0].function
+        # Always expect tool calls - native tool calling is always enabled
+        if not response.tool_calls:
+            raise InvalidAgentResponseError("Assistant did not use a tool")
+        assistant_reply_dict["use_tool"] = response.tool_calls[0].function
 
         parsed_response = OneShotAgentActionProposal.model_validate(
             assistant_reply_dict
         )
-        parsed_response.raw_message = response.copy()
+        parsed_response.raw_message = response.model_copy()
         return parsed_response

classic/original_autogpt/autogpt/app/config.py

@@ -9,18 +9,20 @@ from pathlib import Path
 from typing import Optional, Union
 
 from forge.config.base import BaseConfig
-from forge.llm.providers import CHAT_MODELS, ModelName
+from forge.llm.providers import ModelName
 from forge.llm.providers.openai import OpenAICredentials, OpenAIModelName
 from forge.logging.config import LoggingConfig
 from forge.models.config import Configurable, UserConfigurable
-from pydantic import SecretStr, ValidationInfo, field_validator
+from pydantic import SecretStr
 
 logger = logging.getLogger(__name__)
 
 AZURE_CONFIG_FILE = Path("azure.yaml")
 
 GPT_4_MODEL = OpenAIModelName.GPT4_O
-GPT_3_MODEL = OpenAIModelName.GPT4_O_MINI  # Fallback model for when configured model is unavailable
+GPT_3_MODEL = (
+    OpenAIModelName.GPT4_O_MINI
+)  # Fallback model for when configured model is unavailable
 
 
 class AppConfig(BaseConfig):
@@ -55,9 +57,6 @@ class AppConfig(BaseConfig):
         from_env="SMART_LLM",
     )
     temperature: float = UserConfigurable(default=0, from_env="TEMPERATURE")
-    openai_functions: bool = UserConfigurable(
-        default=False, from_env=lambda: os.getenv("OPENAI_FUNCTIONS", "False") == "True"
-    )
     embedding_model: str = UserConfigurable(
         default="text-embedding-3-small", from_env="EMBEDDING_MODEL"
     )
@@ -90,16 +89,6 @@ class AppConfig(BaseConfig):
         default=AZURE_CONFIG_FILE, from_env="AZURE_CONFIG_FILE"
     )
 
-    @field_validator("openai_functions")
-    def validate_openai_functions(cls, value: bool, info: ValidationInfo):
-        if value:
-            smart_llm = info.data["smart_llm"]
-            assert CHAT_MODELS[smart_llm].has_function_call_api, (
-                f"Model {smart_llm} does not support tool calling. "
-                "Please disable OPENAI_FUNCTIONS or choose a suitable model."
-            )
-        return value
-
 
 class ConfigBuilder(Configurable[AppConfig]):
     default_settings = AppConfig()

classic/original_autogpt/tests/conftest.py

@@ -93,7 +93,6 @@ def agent(
             fast_llm=config.fast_llm,
             smart_llm=config.smart_llm,
             allow_fs_access=not config.restrict_to_workspace,
-            use_functions_api=config.openai_functions,
         ),
         history=Agent.default_settings.history.model_copy(deep=True),
     )

classic/original_autogpt/tests/integration/agent_factory.py

@@ -26,7 +26,6 @@ def dummy_agent(config: AppConfig, llm_provider: MultiProvider):
         config=AgentConfiguration(
             fast_llm=config.fast_llm,
             smart_llm=config.smart_llm,
-            use_functions_api=config.openai_functions,
         ),
         history=Agent.default_settings.history.model_copy(deep=True),
     )