How to make Pydantic AI output model intent as text before tool call (like Claude Desktop)?

[gscfwid]

Question

I’m trying to achieve a behavior similar to Claude Desktop, where the model first declares its intent in plain text before making a tool call. For example, Claude might say: “I will search xx website for you,” and then proceed to call the tool.

In my experience with Pydantic AI, when using tool calls, the model seems to go straight to the tool invocation without outputting any text content that explains its intent or reasoning beforehand.

Is there a recommended way to have Pydantic AI output a text message (e.g., model reasoning or intent) before executing a tool call?

• Is this possible with the current API or configuration?
• Are there any workarounds or best practices for this use case?
• If not, is this feature planned or being discussed?

Additional Context

Pydantic AI version: 0.6.2
Python version: 3.12

[pydanticai-bot]

PydanticAI Github Bot Found 1 issues similar to this one:

1. "best practice that allows an agent to output a natural dialogue text and then make tool calls？ #2411" (90% similar)

[DouweM]

@gscfwid That's going to come down to how you prompt the model, or which model you use -- Anthropic tends to always pre-announce its tool calls.

Pydantic AI only suppresses that when using an output_type, as that forces the model to only output JSON matching the schema, so no regular text.

If you need an output type and tool call pre-announcements, you can add a special tool for the model to call and prompt it to always use that ahead of other tool calls.

The easiest way to get access to all the streaming text, tool calls, and tool results, is to use run with an event_stream_handler: https://ai.pydantic.dev/agents/#streaming-all-events

[zia1138]

@DouweM I was wondering the same thing as @gscfwid. Is it possible to have pydantic-ai return after a tool call, append to messages, and then prompt the model to produce text? Maybe using agent.iter?

[gscfwid]

Hi @DouweM, thank you for the quick reply and the sample code. I’ve tested it on my side:

– When the model immediately issues a tool call, the event stream behaves as expected.
– When the model first writes a short natural-language acknowledgement and then calls the tool, I never see any FunctionCall/tool-related event.

A minimal trace looks like this:

1. PartStartEvent(index=0, part=TextPart(content='I will'))
2. FinalResultEvent(tool_name=None, tool_call_id=None)
3. Final output: I will…

There are only these three events, and the final output is just the intent statement from the first event—no actual tool call occurs.

I’m using OpenAI-compatible providers (Moonshot and OpenRouter). If I add a prompt that forces the model to skip the preamble and call the tool first, everything works, which makes me think the event_stream_handler may be mis-parsing responses that begin with free-text.

Do you have any pointers on debugging this or known workarounds? I can share full logs or a minimal repro script if that helps. Thanks again for your help!

[gscfwid]

I’ve conducted some additional tests using Gemini Flash 2.5.

With my agent built using the MCP tool, it generally takes about five tool calls before arriving at the final output.

When running with ‎run_stream and providing an ‎event_stream_handler, I noticed the model often returns the final output prematurely during the tool call process. In contrast, using the ‎run method usually yields the complete final result as expected.

[aristide1997]

@gscfwid i think that’s a known behaviour with run_stream(). I’d suggest using iter() and stream with deltas instead.

[gscfwid]

@aristide1997 Thank you very much! you solved my problem 100%!