WORTH TESTING · 25 OCTOBER 2026
Google's Agent102 lab: build a self-improving agent (not the '2-hour course' the post promises)
A re-upload by @0xCodila of material from Annie Wang, a software engineer at Google Cloud. Verdict: a solid, free hands-on lab rather than a full course; intermediate, coding needed (terminal, JSON and a little Python) plus a Google Cloud project with billing. It has ten things to test in the part we've seen.
Try this
- Open Google Cloud Shell so you start from a clean environment. 14:15
- Copy the lab code: git clone github.com 14:26
- Run ./setup_project.sh to set up your project and billing. 14:38
- Install uv, then run ./setup_codelab.sh. 14:55
- Launch the agent's web interface with uv run bash table.sh. 15:47
- Open host/instruction_current.txt to see the instructions the agent is working from. 17:15
- Score the agent against a 'golden' test set: uv run adk eval --eval_config eval/eval_config.json (using parties_val.evalset.json). 17:51
- Run the same checks as automated tests with uv run pytest tests/test_host.py. 18:48
- Add your own AI-judge metric in eval/eval_config.json (the lab's example is called 'everyone_ate'). 19:28
- Read loop/flywheel.py: a host, a judge and a proposer loop until the score passes a threshold. 20:06
Timestamps point to the moment in the source video.
Cost
The lab is free; you need a Google Cloud project with billing.
How we knowSteps to try, not a claim that we independently tested every result. Follow the linked sources and check current prices. Our methods & sources ↗