An agent as a tool

Sometimes you want another agent's answer, not another agent's conversation. Turn it into a tool and it reports back.

writer = Agent(
    name="Writer",
    instructions="Write one polite sentence.",
    model=PretendModel(["We are sorry about the delay and are looking into it."]),
)

write_reply = writer.as_tool(
    tool_name="write_reply",
    tool_description="Write a polite reply to a customer.",
)

as_tool gives you an ordinary tool. The name and description are yours to write, and they matter exactly as much as they did in lesson 2, because this is what the calling agent reads when it decides.

result = await Runner.run(support, "The order is late, tell them something")
print("tool the agent had:", [t.name for t in support.tools])
print("answer:            ", result.final_output)

Support stayed in charge the whole time. Writer ran, returned its sentence, and Support used it and wrote the final answer. Compare that with lesson 13, where Billing answered the customer itself.

What it is good for

• A second opinion. Ask a checker agent whether a draft is acceptable, then act on the answer.
• A narrow skill. Translation, summarising, formatting. Things with an input and an output.
• Fan out. One agent can call three others and combine what they say, which a handoff cannot do.