Autogen Tool Use And Function Calling

Operator wraps `def get_weather(city: str) -> str` as a FunctionTool and the model consistently passes city as an object instead of a string.

FunctionTool derives the input schema from the callable's type hints — explicit annotations are required for stable schemas. If the model is passing the wrong shape, tighten the schema (use Pydantic models for nested inputs) and rely on FunctionTool's validation rather than coercing in the function…

FunctionTool implementing send_invoice raises ConnectionError mid-call. Agent has reflect_on_tool_use=True.

AutoGen catches the tool exception and emits a tool result with the error message into the next model context so the model can correct or fall back. Verify the error message is observable (logged + surfaced) and does not include secrets. Do NOT swallow the exception silently — that produces a tool …

Agent has two tools: search_local_docs (project files) and search_web (internet). Both have one-line descriptions; the model routinely chooses search_web for local questions.

Tool descriptions must disambiguate when/why to pick one tool over the other. Write descriptions like 'search_local_docs: search the project's indexed code and markdown only — prefer this when the user references files in the repo'. AutoGen passes the description to the model verbatim; clarity at t…