FastA2A server equivalent to MCPServer

[ArthurJ]

Description

Implementing an MCP server (with FastMCP) and connecting an agent to it (with MCPServerSSE) is easy.

With Pydantic AI, is also easy to make an agent available through A2A protocol (with FastA2A/ .to_a2a method), but connecting an agent to another using the A2A protocol is not as easy.

It would be nice to have a way to do that connection as easily as it is with the MCP, where we can simply give the server in the toolset argument. It could* be on tool/toolset arguments, or as a separated argument (remote_agents maybe?).

References

No response

[DouweM]

@Kludex Does this sound feasible?

[Kludex]

@DouweM Do we have an API for the handsoffs agent in mind already? I think this would fit there.

[DouweM]

@Kludex There's work being done on that in #2589 -- the API will likely still need some tweaking before it's ready to be merged. That allows handoffs to AbstractAgent instances. Are you thinking of an A2AAgent?

[Kludex]

@Kludex There's work being done on that in #2589 -- the API will likely still need some tweaking before it's ready to be merged. That allows handoffs to AbstractAgent instances. Are you thinking of an A2AAgent?

Hmmm... I guess something like that could do?

from pydantic_ai import Agent

agent = Agent('openai:gpt-4.1', handoffs=[A2Agent("https://a2a.pydantic.dev/")])

[ArthurJ]

Also:

from pydantic_ai import Agent

secondary_agent = Agent('openai:gpt-4.1').to_a2a(...)

main_agent = Agent('openai:gpt-4.1', handoffs=[secondary_agent])

[DouweM]

@ArthurJ In that case, to_a2a should not be necessary as handoffs would directly take the secondary agent.