Agent Evaluation
Build a quality flywheel that can distinguish a real improvement from a lucky run.
Define the evaluation contract
Name the target behavior, users, risks, baseline, candidate, environment, stochastic settings, and decision threshold. Start with a few realistic cases, including a boundary or failure case. Split trigger-query optimization into a fixed training set and held-out validation set.
Use eval-schema.md for cases, assertions, timing, and result records.
Run isolated comparisons
- Snapshot the baseline before changing the candidate.
- Run baseline and candidate on identical inputs in fresh contexts with no leaked expected answer or previous trace.
- Capture final artifacts, public transcript/tool summaries, duration, token or request cost, and failures.
- Grade deterministic assertions first; use a blinded rubric or human review for qualities that cannot be measured mechanically.
- Repeat stochastic cases enough to expose variance. Do not hide flakiness by dropping inconvenient runs.
Measure task success, instruction adherence, tool selection and arguments, trajectory efficiency, grounding, safety, output quality, latency, and cost only when relevant. A single aggregate score must not hide a release-blocking metric.
Analyze and iterate
Cluster repeated failures by cause, change one owning layer, rerun the affected cases, then run the regression set. Compare candidate against baseline and reject improvements that regress a protected metric beyond its tolerance.
Use writing-skills for skill-specific authoring and release-engineering for
production promotion. Never claim a score that was not read from an actual
result artifact.
Completion condition
Cases, environment, baseline, candidate, artifacts, graders, costs, and limits are reproducible; the decision follows predefined thresholds rather than a post-hoc interpretation of the preferred result.
