Nous details the review costs behind its 1,393-agent Hermes refactor

Nous Research has published a September account of a large Hermes cleanup that is more instructive than its headline labor-savings comparison. Teknium says the main run dispatched 1,393 subagents over about nineteen active hours, peaking at 218 concurrent workers. The refactor began September 2 and merged September 4 after a restart, a continuation session and two rounds of community review. This is analysis of that earlier event, not a September 23 release.

The account describes 36 non-overlapping assignments, worker git worktrees and delegation reaching three levels below the coordinating agent. Tool schemas and command help were among the interfaces compared with a frozen baseline. Inference ran remotely while an i7 desktop with 64 GB of memory hosted coordination and tool subprocesses. An authentication failure stopped the first run; a separate session prepared a handoff so workers could resume from their saved work.

The awkward details are the useful ones. Reviewers caught removed public names still needed by external plugins, and a rewrite of exception-handling calls introduced regressions at roughly 65 sites. Existing tests had missed those changes. Fixes were required both before and after the merge. The retrospective presents a substantial automation result, but not an autonomous process that made review unnecessary.

The facts

- The author reports 1,393 dispatched subagents and a peak of 218 running concurrently. - Non-test Python fell from 1,063,826 lines to 698,363, a reported reduction of 34.4%. - The reported model cost was about $19,300 for the main run and roughly $25,000 including follow-ups, excluding human review. - A simulated lookup study of 4,000 symbols reduced average returned tokens from 2,218 to 993. - Median lookup tokens increased, and the study did not measure engineering-task completion. - The account says import dependencies increased and some entry points became slower to import.

Why it matters

The result separates three claims often bundled together: code became smaller, some lookup operations became cheaper, and a team estimated substantial avoided manual work. Only the first two have concrete measurements in this account. Staffing estimates are counterfactual comparisons, not audited savings. Operational recovery and review consumed real effort even though model execution handled much of the editing.

Current

Checked September 23, 2026. The inspected release baseline is v2026.9.21. The primary source is maintainer retrospective. Tagged release availability and development-source state are separate: a release baseline does not establish that a proposed setting or newly merged feature is in an installed package.

Evidence

Primary: Teknium / Nous Research — refactoring Hermes (https://nousresearch.com/refactoring-hermes-with-1393-agents). Context and implementation: NousResearch/hermes-agent PR #102117 (https://github.com/NousResearch/hermes-agent/pull/102117); Teknium — refactor benchmark data (https://gist.github.com/teknium1/a7adb797243d6355c76abc9cae88838b). The linked source text was inspected; reported experiments and test results remain attributed to their authors.

Operator take

Use the case study as a design reference for bounded parallel work: non-overlapping ownership, saved commits, stable-interface checks and explicit restart handoffs. Do not use its maximum worker count as a capacity recommendation. The same account describes duplicated language-server processes consuming gigabytes and subsequent resource-sharing repairs. A worthwhile replication would measure downstream task accuracy and latency alongside source size, imports and review hours.

Caveat

This is a maintainer-authored retrospective, surfaced in HN discussion on September 17, about a September 4 merge. Model costs and staffing comparisons are author estimates. Lookup efficiency is not a demonstrated improvement in task success, and human review costs are excluded.