CoworkOS

CoworkOS is an open-source cowork harness from Renmin University’s RUC-NLPIR and collaborators for long-horizon, human-in-the-loop productivity work (repo, project page, arXiv:2611.03178). A three-layer runtime separates Main Agent (plan / delegate / accept), Coordinator Runtime (dispatch and lifecycle), and optional Subagents. Hierarchical workspaces, semantic memory, and revision-aware candidate edits stay on your machine so people and agents can co-edit the same documents, presentations, CAD, or Blender scenes.

Ship surface includes the coworkos-app npm package (desktop install targets Apple Silicon macOS 14+). The paper reports five recursive self-improvement rounds where GLM-5.3 drove post-training of Qwen3.6-35B-A3B, lifting an eight-benchmark average from 33.19 to 42.95. The public repo landed 2026-11-03, with X / HuggingPapers discussion continuing into 2026-11-04.

Problem

Generic chat agents often lose state across long jobs that need human approval, recovery, and shared artifacts. CoworkOS separates strategy from execution and keeps workspaces, logs, and editable artifacts so humans and agents can continue the same project across hours or days.

Differentiation

• Explicit three-layer split: models decide; the runtime coordinates and persists.
• First-class shared apps (docs, slides, CAD, Blender) with revision-aware candidate changes—not only CLI tool calls.
• Paper uses the harness itself as an RSI vehicle (stronger model post-trains a weaker one across rounds).

Limits

• Desktop install docs currently emphasize Apple Silicon macOS; treat other platforms as TBD.
• Benchmark gains are author-reported; reproduce with matching models, data, and harness versions.
• Many products use “cowork” in marketing—cite RUC-NLPIR/CoworkOS or the paper title.

Who it is for

Individuals and small teams doing long-horizon co-editing (research writing, delivery, CAD/3D, monitored loops), plus labs studying agent harnesses / RSI.

Getting started

1. Read the project page and arXiv:2611.03178.
2. Install per the repo: npx coworkos-app@latest desktop install (Node.js ≥22.19.0), then coworkos init / coworkos doctor.
3. Start with a long task that has clear acceptance checks (e.g. multi-file report + human review gates) and inspect Coordinator logs plus workspace revisions.