Google's FlowAgent fixes failing CI tests in 9.85 min, developers applied 28,554 fixes in production
rohanpaul_ai · x · 2026-11-01
- Google's paper introduces FlowAgent, an internal agent that repairs failing tests in the pre-submit CI phase — key finding: latency decides whether developers actually adopt the fix.
- Existing repair agents mostly work offline post-merge; failing CI tests before merge pull developers out of flow. At Google, the median time from failure notification to a developer's first code edit is 27.48 minutes — the golden window for automated repair.
- FlowAgent catches each failure and runs an edit→build→rerun loop (ReAct-style generate-and-validate) with pre/post-execution abstention filters; verified fixes land in code review with a root-cause note.
- Results: experts judged 67.18% of fixes correct on 195 real failures; Google-wide it suggested 295,508 changes, developers previewed 65,069 and applied 28,554. Median run: 9.85 minutes, vs the 27.48-minute window — yet many fixes still arrive too late.
Related event: Google's FlowAgent Auto-Fixes Failing CI Tests(2 posts)→
More from coding & agent
- ai-engineering-hub tutorials repo surges past 39.7k GitHub stars — patchy631 · 2026-11-01
- AiSOC: open-source AI security operations center with LLM-agent triage hits 5.8k stars — beenuar · 2026-11-01
- HKUDS' open-source AI tutor DeepTutor 1.7 ships topic-to-video lessons and in-note Q&A — YiMaTweets · 2026-11-01
- Devs find argon/carbon split complex coding tasks into "task waves" for higher success rates — cgarciae88 · 2026-11-01
- Same model, 17 points apart: OpenCode hits 82% on 21 MCP servers vs Claude Code's 65% — alejxb · 2026-11-01
- mitsuhiko: an Opus 5.5 agent could re-implement all of CPython in two weeks, no humans — mitsuhiko · 2026-11-01