J.P. Morgan OA 面试真题解析：实现 ReAct Agent、工具调用与最终答案解析

Assume these are correctly implemented and available:

llm(prompt, system) -> str  # calls an LLM, returns text
parse_action(text) -> (name, input) | None
parse_final_answer(text) -> str | None
build_system_prompt() -> str
TOOLS: dict[str, Callable]  # {"calculator": fn, "lookup": fn}

Implement the following class:

class ReactAgent:
    def __init__(self, max_iterations: int = 10):
        self.max_iterations = max_iterations
        self.scratchpad: list[str] = []

    def _build_prompt(self, question: str) -> str:
        history = "\n".join(self.scratchpad)
        return f"Question: {question}\n\n{history}"

    def run(self, question: str) -> str:
        system = build_system_prompt()
        for step in range(self.max_iterations):
            prompt = self._build_prompt(question)
            response = llm(prompt, system=system)
            self.scratchpad.append(response)

            action = parse_action(response)
            if action is not None:
                tool_name, tool_input = action
                observation = TOOLS[tool_name](tool_input)
                continue

            final = parse_final_answer(response)
            if final:
                return final

            self.scratchpad.append("Observation: Could not parse an action or final answer. Please try again.")

        return "Agent stopped: max iterations reached."

if __name__ == "__main__":
    agent = ReactAgent()
    answer = agent.run("What year was Python created, and what is 123 * 456?")

Your task is to complete the agent logic so it can iteratively call tools, maintain a scratchpad, detect a final answer, and stop gracefully when the maximum number of iterations is reached.

这道题考察的是 ReAct 风格智能体的核心执行流程：每轮先把问题和历史 scratchpad 组装成 prompt，交给 LLM 生成响应；如果响应中能解析出工具调用，就从 TOOLS 字典里找到对应函数并执行，然后继续下一轮；如果能解析出最终答案，就立即返回；如果两者都不是，则把“无法解析”的观察信息写回 scratchpad，让模型重新尝试。题目重点在于状态管理、工具路由、循环控制和失败兜底，适合用字符串解析、字典映射与迭代循环来实现。