今天則補上最後一塊拼圖：Tools。讓整個 agent-brain 真正動起來。

BaseTool 抽象介面

我先定義了一個最小的 BaseTool 介面，包含三個東東： (反正就按照 MCP server 的概念，之後應該也要能 support mcp server)

1. name：工具的唯一名稱，會被 Reasoning state 引用。

2. description：文字描述，能直接塞進 prompt，讓 LLM 知道這個工具用途。

3. input_schema：輸入格式，用類似 JSON Schema 的結構來定義。

   • 這樣能保證 prompt 時清楚列出參數格式。
   • 之後也能搭配驗證器（Pydantic or jsonschema）做更嚴謹的檢查。

程式骨架如下：

from abc import ABC, abstractmethod

class BaseTool(ABC):
    def __init__(self, name: str, description: str, input_schema: dict):
        self.name = name
        self.description = description
        self.input_schema = input_schema

    def to_dict(self) -> dict:
        return {
            self.name: {
                "description": self.description,
                "input_schema": self.input_schema,
            }
        }

    @abstractmethod
    async def execute(self, **kwargs) -> str: ...

• to_dict()：序列化成統一格式，方便 Memory 拿去丟進 system prompt。
• execute()：真正執行工具邏輯，並回傳結果字串。

最小範例：加法工具

接著我寫了一個最小的工具 AddTwoNumbersTool，用來示範整個流程。

class AddTwoNumbersTool(BaseTool):
    def __init__(self) -> None:
        super().__init__(
            name="add_two_numbers",
            description="Add two numbers.",
            input_schema={
                "type": "object",
                "properties": {
                    "a": {"type": "number", "description": "The first number."},
                    "b": {"type": "number", "description": "The second number."},
                },
                "required": ["a", "b"],
            },
        )

    async def execute(self, **kwargs) -> str:
        a = kwargs.get("a", 0)
        b = kwargs.get("b", 0)
        return f"tool result: {a} + {b} = {a + b}"

Demo

model: gpt-4.1-mini

import asyncio

from agent_brain import BaseTool, Brain


class AddTwoNumbersTool(BaseTool):
    def __init__(self) -> None:
        super().__init__(
            name="add_two_numbers",
            description="Add two numbers.",
            input_schema={
                "type": "object",
                "properties": {
                    "a": {"type": "number", "description": "The first number."},
                    "b": {"type": "number", "description": "The second number."},
                },
                "required": ["a", "b"],
            },
        )

    async def execute(self, **kwargs) -> str:
        a = kwargs.get("a", 0)
        b = kwargs.get("b", 0)
        return f"tool result: {a} + {b} = {a + b}"


brain = Brain(net="ReAct", memory="Messages", tools=[AddTwoNumbersTool()])
task = "What is the value of 12345 + 67890 + 999999?"


async def main() -> None:
    async for chunk in brain.answer(task):
        print(chunk, end="", flush=True)


asyncio.run(main())

來看看會長怎樣吧

@@

to-do:

要可以 support mcp server，e.g., user 只要提供 url