Introduction
Kickstart evaluation of AI agents with Google ADK
()
1. Evaluate a Working AI Agent
Project kickoff: Running a procurement agent
()
When “correct” is dangerous: A policy violation demo
()
Why final answers lie for agents
()
Paths, not strings: Execution trajectories
()
2. Make the Agent Eval‑Ready
ADK project structure and agent.yaml
()
Binding tools to the agent
()
Schema as a contract: Pydantic enforcement
()
First eval: Did the agent call the right tool?
()
3. Expose Logic Failures
Trap mocks: Forcing the agent to reason
()
Anatomy of a trace: Thought, action, observation
()
Visual debugging with ADK Trace View
()
4. Formalizing Evaluation
Capturing golden traces
()
Trajectory matching rules
()
Testing memory: Context persistence
()
Organizing EvalSets for scale
()
5. Scaling Evaluation with Metrics
Running headless eval batches
()
Interpreting scores: Trajectory vs. semantic
()
Pass@k and non‑determinism
()
Reliability vs luck
()
6. Judges, Guardrails, and Production Readiness
LLM‑as‑a‑judge: Custom rubrics
()
Groundedness and faithfulness checks
()
Safe refusal via negative logic
()
Regression gates in CI/CD
()
7. Synthesis and Next Steps
Debugging playbook: Prompt vs. tool vs. model
()
From vibe checks to verifiable agents
()
Congratulations and keep going
()