Claude Code와 LangChain Skills를 활용한 AI 에이전트 개발 및 평가 가이드

claude-code "Make me a deep agent for research that uses Tavily to search the web. It should be able to spin up a subagent for focused topic research, while leaving the main agent to coordinate tasks."

def trajectory_match_evaluator(run, example):
    actual_trajectory = [t['task'] for t in run.outputs['trajectory']]
    expected_trajectory = example.outputs['expected_trajectory']
    # 일치율 계산 로직
    score = len(set(actual_trajectory) & set(expected_trajectory)) / len(expected_trajectory)
    return {"score": score, "key": "trajectory_match"}