Is Pydantic AI suitable for multi-agent systems with complex hand-offs?

All answers to this question.

You can definitely do multi-agent systems with Pydantic AI by using agents as tools. Basically, Agent A has a tool called "consult_specialist" which internally triggers a run of Agent B. Because everything is typed, Agent A knows exactly what format of data it will get back from Agent B. We use this for a coding assistant where a "Reviewer Agent" calls a "Linter Agent." It’s much more predictable than the autonomous "swarm" behavior you see in some other libraries, which can sometimes loop forever or hallucinate transitions.

   Answered 2025-08-20 by Heather Martinez

How do you handle the shared memory or conversation history between those agents in Pydantic AI? If Agent B needs to know what the user said three turns ago to Agent A, is that easy to pass along?

   Answered 2025-08-25 by Jeffrey Moore

• You usually pass the message_history explicitly. It gives you full control. You can decide exactly which parts of the history Agent B needs to see, which actually saves on token costs because you aren't sending a massive, irrelevant context window to every sub-agent.

     Commented 2025-08-28 by Daniel King

For very complex, non-linear flows, you might want to look at pydantic-graph, which is their specialized module for state machines. It's built to handle those "circular" logic paths.

   Answered 2025-09-05 by Larry Taylor

• Good shout on the graph module! It really fills the gap for those of us who need more than just simple linear delegation.

     Commented 2025-09-08 by Michelle Young