A practical evaluation plan for agent tasks, tools, policies, recovery, cost, latency, and human review.
Define the task contract
State the starting context, permitted tools, expected result, stop condition, policy constraints, and acceptable routes to the answer. Separate the quality of the final output from the quality of the path because an agent can occasionally reach a plausible answer through unsafe or wasteful actions.
Build a representative evaluation set
Include routine cases, ambiguous inputs, missing data, conflicting sources, tool errors, permission failures, adversarial instructions, and situations that require escalation. Preserve real failure examples after removing sensitive data so the set grows with operational learning.
Use layered graders
Combine deterministic assertions for schemas, permissions, required citations, tool arguments, and state transitions with rubric-based review for usefulness and judgment. Human review is especially important while defining the rubric and for high-consequence edge cases.
Turn evaluation into a release gate
Track results by workflow version, model, prompt, tool set, and policy configuration. Set thresholds for critical failures, not only an average score. Re-run the suite when any dependency changes and monitor production traces for new cases that the offline set did not represent.
References
This is an internal evaluation framework informed by OpenAI’s public eval guidance. Appropriate tests depend on the workflow and consequence of failure.