Hermes refactors its million-line codebase

Teknium asked his regular Hermes agent to simplify the Hermes repository, split oversized files and remove at least thirty percent of its source lines. The account says the main run lasted about nineteen active hours and dispatched 1,393 subagents, with as many as 218 running concurrently.

The orchestrator measured the repository and divided it into 36 groups. Workers used isolated Git worktrees, written briefs and explicit interface constraints. The parent coordinated assignments and integration rather than editing source itself. Checks compared behavior with a frozen baseline, including tool schemas and command help output.

The run did not proceed uninterrupted. A provider authentication token expired about fifty minutes in. Existing commits and briefs survived, allowing a separate Hermes session to diagnose the problem and continue. The author describes follow-up work and community review before merging the refactor on September fourth.

The account also records fixes after the merge. The reported reduction was 34.4 percent of non-test Python source. Cost estimates excluded human review, and the staffing comparison was an estimate rather than a measured saving. This is useful evidence for a large cleanup workflow precisely because restart recovery, baseline checks and outside review are part of the outcome. A large subagent count alone does not establish that every change was correct.

Prompt
Original prompt

Source: nousresearch.com · Teknium