Nous Research Refactors Million-Line Codebase with 1393 Subagents in 19 Hours
Nous Research published a blog post detailing a large-scale autonomous agent refactoring experiment: on September 2, Teknium handed cleanup of the Hermes open-source codebase—over 1 million lines of non-test Python code—to their own Hermes Agent. The agent orchestrated 1393 subagents, ran for 19 hours, shrank the codebase by 34.4%, and saved an estimated $2 million in engineering time.
Confirmed
- The task began on September 2, when Teknium assigned the codebase cleanup to the Hermes Agent
- The repository contained over 1 million lines of non-test Python code
- 1393 subagents were dispatched, with 19 hours of total runtime
- The codebase was reduced by 34.4%, saving nearly $2 million at typical engineering hourly rates
Why it matters
- This is a rare publicly documented case of autonomous refactoring of a million-line codebase, demonstrating the feasibility of multi-subagent orchestration on real large-scale engineering tasks
- The cost savings, derived from engineering hourly rates, offer a quantified reference for evaluating agent engineering value
- Built on their own Hermes model and agent framework, it is an empirical demonstration of open-source community model capabilities
2026-09-16 ~ 2026-09-17 · 5 related posts
Primary sources
- [source] Nous Research Used 1,393 Subagents to Cut a 1M-Line Codebase by 34.4% in 19 Hours — NousResearch · 2026-09-16
4 near-duplicate retellings: Teknium · NousResearch · max_paperclips · omarsar0