Google ran Argon overnight on a 200K-line codebase: 40% faster sessions, 2 of 6 fixes shipped
rseroter · x · 2026-10-14
Google's developer blog details an 11-hour unattended overnight run where Gemini 4 Argon tackled a much harder problem than greenfield demos: finding cross-file performance bottlenecks in google/adk-python (205,000 lines, 800 source files, 17,000 unit tests). Using .agents/skills rulefiles and /goal mode, the agent proposed 6 changes; 2 shipped after human review — cutting in-memory session prep time by 40% in long sessions and speeding cold package imports by 21%, while 4 were rejected in review. The post includes the full prompts, skills configs, and review decisions, plus how to set up the same overnight workflow on your own codebase.
More from coding & agent
- A 'council' of AI agents answers in parallel, then votes via chair — johnlindquist · 2026-10-14
- Dev Flags Agent Flaw: pkill Matches Its Own Shell and Kills the Session — tristanbob · 2026-10-14
- Mysterious Blackbox plugin lands in Hermes Agent catalog to guard AI agents from compromise — Teknium · 2026-10-14
- Obsidian Starter Kit v5 ships Claude Code slash commands like /checkin and /closeday — dSebastien · 2026-10-14
- Modder plugs a local 2B decision model into Halo CE as a split-screen AI co-op teammate — CurtisSawtooth · 2026-10-14
- Agentic software factory workshop demos a queue-driven agent system built live in Codex — hugobowne · 2026-10-14