MCP tools appear to be added twice

[wreed4]

I'm still tracking this down, but as far as I can tell, by the time we get to the code in pydantic_ai/_agent_graph.py function _prepare_request_parameters, we get to this line

--> 241     await asyncio.gather(
    242         *map(add_tool, ctx.deps.function_tools.values()),
    243         *map(add_mcp_server_tools, ctx.deps.mcp_servers),
    244     )

And the MCP tools are already added to ctx.deps.function_tools in the actor.get_agentlet function in orchestrator.py line 173 in Orchestrator.get_once. So they get added twice and sent to the llm twice, which causes a confict because names are re-used.

I'll keep digging in a bit more and see if I can see why this is happening, but I'm brand new to this codebase and this project.. this is my first attempt at using it :(.

[wreed4]

I anticipate this is because pydantic.Agent added (maybe?) auto discovery of mcp tools and this is also doing discovery of MCP tools.

[wreed4]

Yes... commenting out this code seems to have fixed it potentially. Though I have no idea what else this could cause.. first time looking at this codebase

        unique_marvin_tools = [wrap_tool_errors(tool) for tool in marvin_tool_callables]

        mcp_tool_instances: list[Tool] = []
        # if active_mcp_servers:
        #     orchestrator = marvin.engine.orchestrator.get_current_orchestrator()
        #     mcp_tool_instances = await discover_mcp_tools(
        #         mcp_servers=active_mcp_servers,
        #         orchestrator=orchestrator,
        #     )

        combined_tools: list[Any] = unique_marvin_tools + mcp_tool_instances

from Agent.get_agentlet, in agent.py:163

[zzstoatzz]

thanks for the report @wreed4 ! your diagnosis is correct and the fix is released in 3.1.1