Google ran Gemini 4 Argon for 11 hours on its ADK repo and rejected four of its six changes

The target was google/adk-python, roughly 205,000 lines across nearly 800 source files with nearly 17,000 unit tests. A first two-hour run, given one paragraph of goal and no repository instructions, found real bottlenecks but altered the public JSON schema of ADK's authentication model, missed mock edge cases, and ran only 216 tests. For the second run the team put its maintainer guidelines in the repository's .agents/skills directory and used Antigravity's /goal mode, which keeps the goal across context compactions. That run went 11 hours, read 12 guideline files first, ran all 16,892 tests, and opened a pull request with six optimizations across 32 files. Reviewers shipped two and rejected four, including a change that sped up a synthetic benchmark by 17% but relied on a weak memory reference that could return stale data across sessions.

The merged change touches 22 files and adds 17 test functions. On a holdout benchmark the model never saw, Google reports about 40% less in-memory session preparation overhead in long sessions and 21% faster cold imports. Time from a new process to the first model request stayed flat, at 3.07 s against 3.09 s. The authors also say 25 internal Google test targets failed on an early lazy-import attempt even though every open-source test passed. Their summary: "the agent explores, the maintainer subtracts."

Why it matters: The transferable part is the setup, not the speedup. The same model went from a run that changed a public schema to one that followed repo conventions once the rules were written as files and the goal survived compaction. The holdout benchmark and human review caught what the agent's own tests did not. If you run agents overnight, write your review conventions into the repo, keep a benchmark the agent cannot see, and budget reviewer time for cutting changes. This is one run on Google's own framework with Google's own model and tooling, and the write-up cites no independent replication.