charlieviettq Agent evaluation
What to measure
| Dimension | Examples |
|---|---|
| Task success | End state matches spec (binary or rubric) |
| Tool use | Correct tool, valid args, no spurious calls |
| Safety | No policy violations, no secret leakage |
| Efficiency | Tokens, latency, tool call count |
| Stability | Same input -> consistent outcome across runs |
Workflow
- Define tasks — realistic user intents with clear pass/fail or scored rubric.
- Build dataset — golden set + edge cases (errors, ambiguous input, empty context).
- Run baseline — fixed model/settings; log traces (inputs, tools, outputs).
- Score — automated checks first; human review for ambiguous cases.
- Compare — A/B prompts, models, or tool schemas; report deltas with confidence notes.
- Gate — block release on regression in must-pass tasks.
Automated checks
- Schema validation on tool arguments.
- Assert final answer contains required fields or avoids forbidden content.
- Snapshot tests for deterministic sub-steps where possible.
Human rubric (when needed)
Score 1-5 on: correctness, completeness, tone, safety. Document disagreements.
Anti-patterns
- Eval only on cherry-picked happy paths.
- Changing task and model simultaneously without isolation.
- No trace logs when debugging tool failures.
Output
Summary table: variant | success rate | avg tools | avg latency | notes.