如何科学测评一个 AI Agent？

示例（Level 2）:
"在 2024 年 Nature 上发表最多 AI 论文的作者是谁？
列出他/她 2024 年的前 3 篇论文标题。"
→ 需要: 搜索 → 网页抓取 → 统计 → 排序 → 格式化

class AgentEvaluator:
    def __init__(self, benchmark: list[Task]):
        self.benchmark = benchmark

    def evaluate(self, agent, max_steps=20):
        results = []
        for task in self.benchmark:
            result = self._run_task(agent, task, max_steps)
            results.append(result)

        return {
            "accuracy": sum(r["success"] for r in results) / len(results),
            "avg_steps": np.mean([r["steps"] for r in results]),
            "avg_cost": np.mean([r["cost"] for r in results]),
            "avg_time": np.mean([r["time"] for r in results]),
        }

    def _run_task(self, agent, task, max_steps):
        start = time.time()
        try:
            output = agent.run(task.prompt, max_steps=max_steps)
            success = self._check_answer(output, task.expected)
        except Exception as e:
            success = False
            output = str(e)

        return {
            "success": success,
            "steps": agent.step_count,
            "cost": agent.total_cost,
            "time": time.time() - start,
        }