Nous Research Used 1,393 Subagents to Cut a 1M-Line Codebase by 34.4% in 19 Hours
NousResearch · x · 2026-09-16
Nous Research details how its Hermes Agent autonomously refactored over a million lines of non-test Python: a 19-hour main run dispatching 1,393 subagents (218 concurrent at peak), followed by community review, cut the codebase by 34.4%. Token costs ran $19.3K ($25K with follow-ups), versus an estimated $150K–$1.8M for manual work. Teknium notes the agent had learned his coding standards from daily dogfooding, updating its skills as they fixed bugs together.
More from coding & agent
- skillmem: open-source procedural memory for MCP agents, with hard-won recursion lessons — mrPetrukovich · 2026-09-16
- Running More Than 2 Sub-Agents in Parallel Burns Tokens for Zero Quality Gain — pvncher · 2026-09-16
- Devin Gets a Cloud Mac: Build and Ship iOS/macOS Apps From a Prompt — Rasmic · 2026-09-16
- Self-evolving agent framework Épi hits 120 research records, tightens Klarner's constant — my_cat_can_code · 2026-09-16
- MCP deferred tools explained: progressive disclosure, like a CLI, model calls first to fetch params — eherrerosj · 2026-09-16
- Muse Code Raises Contributor Tier Limits, Resets All Subscription Usage — armand_ruiz · 2026-09-16