Fix: Functions Passed to Agents Are Now Available as Tools

[rholinshead]

Summary

Currently, functions passed to agents do not properly get listed as tools since the async initialize is not (always?) called. To fix, pull the _function_tool_map construction into __init__.

Changes

• Construct _function_tool_map in __init__ in agent.py
• Add functions example to show working usage

Test Plan

uv run scripts/example.py run functions --debug

See our functions listed as tools:

[INFO] 2025-04-07T20:15:13 mcp_agent.mcp_agent_using_functions - Tools available:
{
  "data": {
    "meta": null,
    "nextCursor": null,
    "tools": [
      {
        "name": "add_numbers",
        "description": "\n    Adds two numbers.\n    ",
        "inputSchema": {
          "properties": {
            "a": {
              "title": "A",
              "type": "integer"
            },
            "b": {
              "title": "B",
              "type": "integer"
            }
          },
          "required": [
            "a",
            "b"
          ],
          "title": "add_numbersArguments",
          "type": "object"
        }
      },
      {
        "name": "multiply_numbers",
        "description": "\n    Multiplies two numbers.\n    ",
        "inputSchema": {
          "properties": {
            "a": {
              "title": "A",
              "type": "integer"
            },
            "b": {
              "title": "B",
              "type": "integer"
            }
          },
          "required": [
            "a",
            "b"
          ],
          "title": "multiply_numbersArguments",
          "type": "object"
        }
      },

See them being called as tools:

...
 {
        "finish_reason": "tool_calls",
        "index": 0,
        "logprobs": null,
        "message": {
          "content": null,
          "refusal": null,
          "role": "assistant",
          "annotations": [],
          "audio": null,
          "function_call": null,
          "tool_calls": [
            {
              "id": "call_fp1JQviBV6p4Yfbb2gU6pASf",
              "function": {
                "arguments": "{\"a\":2,\"b\":3}",
                "name": "add_numbers"
              },
              "type": "function"
            }
          ]
        }
      }
...

 {
        "finish_reason": "tool_calls",
        "index": 0,
        "logprobs": null,
        "message": {
          "content": null,
          "refusal": null,
          "role": "assistant",
          "annotations": [],
          "audio": null,
          "function_call": null,
          "tool_calls": [
            {
              "id": "call_P4LwnJKUtYQ8wjKLxMo5JIt7",
              "function": {
                "arguments": "{\"a\":5,\"b\":4}",
                "name": "multiply_numbers"
              },
              "type": "function"
            }
          ]
        }
      }
...

And correct response:

[INFO] 2025-04-07T20:15:16 mcp_agent.mcp_agent_using_functions - Expert math result: The result of adding 2 and 3 is 5. 
When you multiply that result by 4, you get 20.

[saqadri]

Thanks for the fix @rholinshead! I think your fix uncovered another issue which is that initialize shouldn't be calling aenter, but rather an aenter should be calling initialize. That mistake is why this wasn't working as intended.

[saqadri]

This should initialize the agent

[saqadri]

It would be good to change the initialize function to get rid of this, and instead to implement the aenter asynccontextmanager for this class, which a) calls super.aenter, and b) calls initialize.

I think for now initialize can stay in empty and just call super.initialize()

[rholinshead]

Ok, I looked into this and:

• confirmed that implementing __aenter__ and calling initialize with the function tool map construction in initialize (to test) worked as expected, e.g.

    async def __aenter__(self):
        """
        Initialize the agent and connect to the MCP servers.
        NOTE: This method is called automatically when the agent is used as an async context manager.
        """
        if self.initialized:
            return self

        await super().__aenter__()
        await self.initialize()
        return self

    async def initialize(self):
        for function in self.functions:
            tool: FastTool = FastTool.from_function(function)
            self._function_tool_map[tool.name] = tool

But, having the function tool map construction in __init__ IMO makes sense (sounds like we agree). Looks like there is no initialize in any superclass so super().initialize() is invalid. In that case, we'll end up with an empty initialize and then __aenter__ becomes a redundant override.
So, might make sense to just remove __aenter__ and initialize from this class for now and add back in the future if needed?

[rholinshead]

@saqadri lastest version here removes them as redundant. If you'd prefer to have them just to be explicit / to allow future updates to initialize please let me know and I can add them back

[saqadri]

This is a good change to have in the constructor.

[saqadri]

thanks!