Annie’s code lab → https://goo.gle/4qTyNPY
Docs for how to evaluate agents in ADK → https://goo.gle/4t1ToDb
AI agents are software, and like any other software, they need to be tested before deployment. In this episode, Google Cloud’s Annie Wang walks Martin Omander through a practical, hands on workflow for evaluating agents using the Agent Development Kit (ADK).
We move beyond simple "vibes-based" testing and look at how to rigorously validate your agent’s behavior throughout the development lifecycle. You’ll see how to:
1️⃣ Run interactive tests using the ADK web interface during local development.
2️⃣ Create "golden datasets" to benchmark your agent’s performance and catch regressions.
3️⃣ Evaluate logic by measuring "trajectory scores" (did the agent call the right tools in the right order?) rather than just the final text response.
4️⃣ Automate your pipeline by moving from the command line (adk eval) to fully automated CI/CD integration using pytest.
Chapters:
0:00 – Intro
1:48 – Types of agent tests
2:25 – Interactive testing
4:58 – Command line testing
5:32 – CI/CD testing
6:17 – Why trajectory testing?
6:45 – How to deal with varying output?
7:40 – Takeaways
#ServerlessExpeditions #VibeCoding
Speakers: Martin Omander, Annie Wang
Products Mentioned: Agent Development Kit
