SuperAPI AI Directory

agent-tool

Agents & models

View repository

Updated 10-16 07:29

About this skill

Add an Agent Tool ({{ cookiecutter.ai_framework }})

Agent tools live in backend/app/agents/tools/ and are surfaced to the model so it can call them mid-conversation. The assistant is defined in backend/app/agents/.

Steps

  1. Write the tool function in backend/app/agents/tools/<tool_name>.py:

    • Async, fully type-hinted, with a clear docstring — the docstring and signature are what the model sees, so make them precise.
    • Pure logic: take typed args, return a JSON-serializable result. Raise on hard errors; return a structured {"error": ...} for soft failures the model should reason about.
    • Keep secrets/IO behind settings and the service layer; don't inline credentials.
  2. Export it from backend/app/agents/tools/__init__.py (add the import and append to __all__, matching the existing feature-gated blocks).

  1. Register it on the agent in the assistant for the active framework: {%- if cookiecutter.use_pydantic_ai %}
    • app/agents/assistant.py — decorate with @agent.tool (needs RunContext[Deps]) or @agent.tool_plain (no context):
      @agent.tool
      async def my_tool(ctx: RunContext[Deps], query: str) -> dict:
          """One-line description the model reads to decide when to call this."""
          ...
      

{%- elif cookiecutter.use_pydantic_deep %}

  • app/agents/pydantic_deep_assistant.py — add the function to the agent's tool list. {%- elif cookiecutter.use_langchain %}
  • app/agents/langchain_assistant.py — wrap with @tool and add it to the tools=[...] passed to the agent. {%- elif cookiecutter.use_langgraph %}
  • app/agents/langgraph_assistant.py — wrap with @tool and include it in the tools list bound to the graph. {%- elif cookiecutter.use_deepagents %}
  • app/agents/deepagents_assistant.py — add the function to the agent's tool list. {%- endif %}
  1. Prompt guidance (optional but recommended): if the tool should only be used in specific situations, add a sentence to the system prompt in app/agents/prompts.py so the model knows when to reach for it.

  2. Frontend rendering (optional): tool calls render as cards in the chat UI. For a bespoke card, add a renderer under frontend/src/components/chat/tool-results/; otherwise the generic card handles it.

  3. Test it: add backend/tests/test_<tool_name>.py (see existing test_web_search.py / test_chart_tool.py). Tools are plain async functions — test them directly, no agent needed.

Rules

  • The docstring is the contract with the model — keep it accurate and action-oriented.
  • Return small, structured payloads; don't dump huge blobs into the context.
  • Long-running or side-effecting work belongs in a service (and possibly a background task), not inline in the tool.
  • See docs/howto/add-agent-tool.md for a fuller walkthrough.