Evals in CI/CD
Make agent quality a required check. See a GitHub Actions workflow run unit tests and AgentEvaluator on every pull request, block a regression, and stay green when only the cost changes.
Checking your access…
This is a Pro lesson
Sign in to continue. Pro members get every animated lesson, its code and its quiz, across ADK, Agent Platform, Agent QA and CX Agent Studio.
You're signed in, but this lesson is part of Pro
Get every Pro lesson, its code and its quiz for ₹999, once. Lifetime access, including lessons added later.
Unlocking your lesson…
If this doesn't change in a moment, reload the page.
In this module: ADK evaluation
- How ADK evaluates an agent Evalsets, trajectory and response scoring, pass and fail, animated. Pro
- Test files vs evalsets .test.json for unit-style, .evalset.json for integration. Pro
- Trajectory and response criteria tool_trajectory_avg_score and response_match_score. Pro
- LLM-as-judge criteria final_response_match_v2, rubrics, hallucinations, safety. Pro
- Multi-turn and user simulation Letting an AI play the user. Pro
- Environment simulation Testing without real backends. Pro
- Custom metrics Scoring what matters to your product. Pro
- Efficiency metrics Tool calls, model calls, tokens and latency. Pro
- Running evals: web UI, CLI and pytest adk web, adk eval and AgentEvaluator. Pro
- Evals in CI/CD Blocking a merge when quality drops. Pro
- Testing deployed agents Smoke tests against a live endpoint. Pro
- Optimization Improving an agent from eval results. Pro