Give a coding agent repository rules before an unattended optimization run
Added to DistroCool on 2026-10-22 · Discovery date, not a release announcement
Google engineers documented an 11-hour Gemini 4 Argon run on the roughly 205,000-line ADK Python repository. They placed maintainer conventions in .agents/skills files, required benchmarks and regression tests, and asked for an unmerged pull request. Human review accepted two of six proposals; Google reports the shipped changes reduced in-memory session preparation overhead and cold import time in its measured cases.
Why it matters
The most reusable lesson is not the vendor’s performance claim but the control structure: encode repository conventions, demand before-and-after measurements, run the full test suite, and stop at a reviewable pull request. The first unguided run changed a public schema, missed edge cases and ran only nearby tests, showing why a broad goal alone is not enough.
Try this next.
On a disposable branch, add a short SKILL.md that links to your architecture rules, visibility conventions, benchmark command and full CI command. Ask the agent to find one measurable bottleneck, preserve public behavior, add a regression test, and open but not merge a pull request with before-and-after numbers. Reject changes that lack reproducible evidence. Gemini 4 Argon access and API usage may be paid. DistroCool has not independently tested this example.
Open the practical workflow →Source: Google engineering case study — October 13, 2026 ↗