Eleven unattended hours, six proposals, two shipped: the agent explores, the maintainer subtracts

George Weale, Bo Yang and Andrew Biernat (Google Cloud) write up an overnight performance run on google/adk-python: roughly 205,000 lines, nearly 800 source files, about 17,000 unit tests. The interesting part is not the speedup. It is the difference between two runs with the same goal.

Round one got a one-paragraph outcome prompt and no repo instructions. It ran about two hours, wrote a benchmark seven minutes in, and spent 70 minutes measuring and reading before editing. It found real bottlenecks, but it broke maintainer conventions: a hook that subtly changed a public JSON schema, missed mock edge cases, only 216 nearby tests run, no numbers in the PR description.

Round two added two things: the team's maintainer guidelines as .agents/skills files (a short index pointing to focused references loaded on demand), and a goal pinned across context compactions, with an explicit ask for one unmerged PR "the ADK maintainers would approve," before-and-after numbers included. It started at 9:11 PM and ran 11 hours unattended. It read 12 guideline files and ran a baseline across four workloads before touching source; three minutes before stopping on its own, it was still auditing its diff, caught a bug in its own multi-agent cache, and reran all 16,892 tests.

The morning PR proposed six optimizations across 32 files. Review shipped two: filtering in-memory session history before copying (about 40% less pre-model overhead in long sessions; a 158-turn run dropped from 65.6s to 56.1s) and lazy-loading heavy packages (21% faster cold import). Even that second fix needed another pass: the open-source suite was green, yet 25 internal Google test targets failed because they patched the removed module attributes. Review cut the other four: a cache keyed on a weak reference that could serve stale data across sessions, a telemetry check reading a private OpenTelemetry attribute, a readability-costly rewrite that didn't move an independent holdout benchmark the agent never saw, and an index that only helped synthetic replay tests. The final 22-file change landed with the commit note "Argon generated these changes."

Why it matters for code review

This is the overnight case at its most honest: the harness (written-down conventions, a goal that survives compaction, a holdout benchmark, "don't merge") turned two hours of plausible output into eleven hours of reviewable output. Still, two thirds of what came back was cut by a maintainer, and a green public suite still hid 25 internal failures. Unattended runs compound when review is a subtraction step with its own evidence, not a rubber stamp. Compare the same model swinging 35 points on harness alone, an overnight dual-model harness, and the Claude Code team's risk-tiered review.