Agents And Workflows

A Python function is wrapped as a FunctionTool but its parameters lack type hints and the docstring is empty, so the generated tool schema is untyped and the agent calls it with wrong argument types.

Give tool functions precise type hints and a clear docstring (or an explicit Pydantic schema / fn_schema) so FunctionTool generates a correct JSON schema the agent can call reliably. Validate arguments against the schema before executing; untyped tools produce malformed calls.

An integrator uses ReActAgent with an LLM that has strong native tool/function calling, then fights with the ReAct text format being mis-parsed.

Use FunctionAgent (native tool calling) with models that support it for reliable structured tool calls; reserve ReActAgent for models without native function calling where reasoning is expressed as text. Match the agent type to the model's capabilities rather than defaulting blindly.

An agent is given a delete_records tool and, prompted by ambiguous user input plus injected text in a retrieved document, attempts to call it with a broad filter.

Gate destructive/irreversible tools behind confirmation, scoping, or human approval; never let model reasoning (especially influenced by retrieved/untrusted content) trigger broad destructive actions unchecked. Validate tool arguments and require explicit confirmation for high-impact operations.