fix(copilot): use bounded default for agent_mode_max_iterations

Change fixer default from -1 (infinite) to 10 (bounded) for safety.
Update guide to let LLM choose iteration count based on task complexity:
1 for single-step, 3-10 for multi-step, -1 for open-ended orchestration.

autogpt_platform/backend/backend/copilot/sdk/agent_generation_guide.md

@@ -150,10 +150,13 @@ call in a loop until the task is complete:
 1. Create a `SmartDecisionMakerBlock` node
    (ID: `3b191d9f-356f-482d-8238-ba04b6d18381`)
 2. Set `input_default`:
-   - `agent_mode_max_iterations`: `-1` (infinite loop — AI calls tools until
-     done) or a positive number for bounded iterations. Do NOT use `0`
-     (traditional mode) — it requires complex external conversation-history
-     loop wiring that the agent generator does not produce.
+   - `agent_mode_max_iterations`: Choose based on task complexity:
+     - `1` for single-step tool calls (AI picks one tool, calls it, done)
+     - `3`–`10` for multi-step tasks (AI calls tools iteratively)
+     - `-1` for open-ended orchestration (AI loops until it decides it's done)
+     Do NOT use `0` (traditional mode) — it requires complex external
+     conversation-history loop wiring that the agent generator does not
+     produce.
    - `conversation_compaction`: `true` (recommended to avoid context overflow)
    - Optional: `sys_prompt` for extra LLM context about how to orchestrate
 3. Wire the `prompt` input from an `AgentInputBlock` (the user's task)
@@ -169,7 +172,7 @@ call in a loop until the task is complete:
 **Example — Orchestrator calling two sub-agents:**
 - Node 1: `AgentInputBlock` (input_default: `{"name": "task"}`)
 - Node 2: `SmartDecisionMakerBlock` (input_default:
-  `{"agent_mode_max_iterations": -1, "conversation_compaction": true}`)
+  `{"agent_mode_max_iterations": 10, "conversation_compaction": true}`)
 - Node 3: `AgentExecutorBlock` (sub-agent A — set `graph_id`, `graph_version`,
   `input_schema`, `output_schema` from library agent)
 - Node 4: `AgentExecutorBlock` (sub-agent B — same pattern)

autogpt_platform/backend/backend/copilot/tools/agent_generator/fixer.py

@@ -33,7 +33,7 @@ _TEXT_REPLACE_BLOCK_ID = "7e7c87ab-3469-4bcc-9abe-67705091b713"
 
 # Defaults applied to SmartDecisionMakerBlock nodes by the fixer.
 _SDM_DEFAULTS: dict[str, object] = {
-    "agent_mode_max_iterations": -1,
+    "agent_mode_max_iterations": 10,
     "conversation_compaction": True,
     "retry": 3,
     "multiple_tool_calls": False,

autogpt_platform/backend/test/agent_generator/test_smart_decision_maker.py

@@ -136,7 +136,7 @@ class TestFixSmartDecisionMakerBlocks:
         result = fixer.fix_smart_decision_maker_blocks(agent)
 
         defaults = result["nodes"][0]["input_default"]
-        assert defaults["agent_mode_max_iterations"] == -1
+        assert defaults["agent_mode_max_iterations"] == 10
         assert defaults["conversation_compaction"] is True
         assert defaults["retry"] == 3
         assert defaults["multiple_tool_calls"] is False
@@ -220,7 +220,7 @@ class TestFixSmartDecisionMakerBlocks:
         result = fixer.fix_smart_decision_maker_blocks(agent)
 
         assert "input_default" in result["nodes"][0]
-        assert result["nodes"][0]["input_default"]["agent_mode_max_iterations"] == -1
+        assert result["nodes"][0]["input_default"]["agent_mode_max_iterations"] == 10
 
     def test_handles_none_input_default(self):
         """Node with input_default set to None gets a dict created."""
@@ -236,7 +236,7 @@ class TestFixSmartDecisionMakerBlocks:
         result = fixer.fix_smart_decision_maker_blocks(agent)
 
         assert isinstance(result["nodes"][0]["input_default"], dict)
-        assert result["nodes"][0]["input_default"]["agent_mode_max_iterations"] == -1
+        assert result["nodes"][0]["input_default"]["agent_mode_max_iterations"] == 10
 
     def test_treats_none_values_as_missing(self):
         """Explicit None values are overwritten with defaults."""
@@ -258,7 +258,7 @@ class TestFixSmartDecisionMakerBlocks:
         result = fixer.fix_smart_decision_maker_blocks(agent)
 
         defaults = result["nodes"][0]["input_default"]
-        assert defaults["agent_mode_max_iterations"] == -1  # None → default
+        assert defaults["agent_mode_max_iterations"] == 10  # None → default
         assert defaults["conversation_compaction"] is True  # None → default
         assert defaults["retry"] == 3  # kept
         assert defaults["multiple_tool_calls"] is False  # kept
@@ -280,7 +280,7 @@ class TestFixSmartDecisionMakerBlocks:
         # First node: 3 defaults filled (agent_mode was already set)
         assert result["nodes"][0]["input_default"]["agent_mode_max_iterations"] == 3
         # Second node: all 4 defaults filled
-        assert result["nodes"][1]["input_default"]["agent_mode_max_iterations"] == -1
+        assert result["nodes"][1]["input_default"]["agent_mode_max_iterations"] == 10
         assert len(fixer.fixes_applied) == 7  # 3 + 4
 
     def test_registered_in_apply_all_fixes(self):
@@ -294,7 +294,7 @@ class TestFixSmartDecisionMakerBlocks:
         result = fixer.apply_all_fixes(agent)
 
         defaults = result["nodes"][0]["input_default"]
-        assert defaults["agent_mode_max_iterations"] == -1
+        assert defaults["agent_mode_max_iterations"] == 10
         assert any("SmartDecisionMakerBlock" in fix for fix in fixer.fixes_applied)
 
 
@@ -528,7 +528,7 @@ class TestSmartDecisionMakerE2EPipeline:
             n for n in fixed["nodes"] if n["block_id"] == SMART_DECISION_MAKER_BLOCK_ID
         ]
         assert len(sdm_nodes) == 1
-        assert sdm_nodes[0]["input_default"]["agent_mode_max_iterations"] == -1
+        assert sdm_nodes[0]["input_default"]["agent_mode_max_iterations"] == 10
         assert sdm_nodes[0]["input_default"]["conversation_compaction"] is True
 
         # Validate (standalone SDM check)
@@ -557,7 +557,7 @@ class TestSmartDecisionMakerE2EPipeline:
         # Fix fills defaults fine
         fixer = AgentFixer()
         fixed = fixer.apply_all_fixes(agent)
-        assert fixed["nodes"][1]["input_default"]["agent_mode_max_iterations"] == -1
+        assert fixed["nodes"][1]["input_default"]["agent_mode_max_iterations"] == 10
 
         # Validate catches missing tools
         validator = AgentValidator()