Agent evaluation checklist
Datasets, graders, trajectory and state checks, and release gates.
QADeveloperProduct
Dataset
- About 50 real or realistic cases to start
- Common, ambiguous, out of scope, hostile, injection and policy-edge cases
- Tag by scenario; keep a holdout
- Every production failure becomes a case
What to grade
| Dimension | Check with |
|---|---|
| Final response | Code, then a calibrated judge |
| Trajectory | Expected path or rules |
| Tool use | Code: tool, arguments, call count |
| State changes | Code: query the system after the run |
| Whole conversation | Judge over the transcript plus human samples |
Judges
- One criterion each, written as pass or fail
- Give the evidence: policy, tool results, transcript
- Label 50 to 100 cases with people first and measure agreement
- Randomize order and test for length bias
- Re-calibrate when the judge, rubric or agent changes
Simulated users
- Personas with goals and constraints, including difficult ones
- Limit turns and record everything
- Compare with real transcripts; do not trust them alone
Release gate
- Fail on any safety case
- Fail on a drop beyond a threshold set from your own baseline
- Fail when a fixed bug returns
- Run offline before release and sample online after