| name | agent-evaluation |
| description | methodologies for assessing agent capabilities, tool-use trajectories, and final response quality. Use this to implement automated testing and human-in-the-loop validation for agents. |
Agent Evaluation Frameworks
Goal
Implement a multi-layered evaluation strategy that assesses not just the final answer, but the reasoning process (trajectory) and core capabilities of the agent.
1. Evaluating Trajectory (The "How")
- Definition: Assessing the sequence of steps (thoughts, tool calls, observations) the agent took to reach a conclusion.
- Metrics:
- Exact Match: The agent's path perfectly mirrors the ideal reference trajectory (rigid).
- In-Order Match: The agent completed the core steps in the correct sequence, ignoring harmless extra steps (flexible).
- Any-Order Match: The agent performed all necessary actions, regardless of sequence.
- Tool Precision/Recall: Did the agent call relevant tools and avoid irrelevant ones?.
2. Evaluating Final Response (The "What")
- Definition: Assessing the quality, relevance, and correctness of the final output provided to the user.
- Mechanism:
- Golden Datasets: Compare output against manually curated "correct" answers.
- Autoraters (LLM-as-a-Judge): Use a strong model to grade the response against specific criteria (e.g., "Is the tone helpful?", "Is the answer grounded in the context?").
3. Assessing Capabilities
- Benchmarks: Use standard benchmarks to test fundamental skills before deployment:
- Tool Calling: Berkeley Function-Calling Leaderboard (BFCL) to test tool selection accuracy.
- Planning: PlanBench to assess reasoning and multi-step logic.
4. Human-in-the-Loop (HITL)
- Purpose: Calibrate automated metrics and assess subjective qualities like "creativity" or "nuance" that machines miss.
- Methods:
- Direct Assessment: Experts score performance on specific tasks.
- A/B Testing: Compare the new agent version against the old one in production.