Start with a measurable outcome
For a support agent, success might require the correct order, an authorized lookup, a supported answer, and no unauthorized refund. A fluent final answer is insufficient. Define deterministic checks for IDs and required fields, plus a rubric for answer quality. This is an application evaluation design, not an automatic guarantee supplied by the runtime.
Debug behavior before scaling tests
OpenAI's workflow evaluation guidance starts with traces to inspect model calls, tools, guardrails, and handoffs. Its trace-grading walkthrough describes SDK apps and existing Builder workflows. Do not assume that every managed Agents API event is automatically the same trace-grading resource; use the observability path documented for your runtime.
Build a representative dataset
Include normal tasks, missing information, ambiguous requests, tool errors, and adversarial content. Remove sensitive customer data or use authorized test fixtures. Keep a held-out set to avoid optimizing only for examples used while editing prompts. Record the configuration and tool versions for each run.
Combine checks with human judgment
Use exact checks for forbidden actions and required outputs, and rubric-based grading for less structured answers. Review disagreements between graders and people. Measure completion, unsafe actions, latency, and cost separately; an average score can hide a critical safety regression.
Gate changes on evidence
Run baseline and candidate configurations against the same cases. Investigate regressions before broad rollout and retain representative failure traces. Repeat trials where model variation matters. Use datasets and eval runs for repeatability rather than relying on one convincing demo.
Read the official reference
Check the source for current API fields, account requirements, and service limits.
OpenAI: agent-evals