# CoworkOS Explained: An Open-Source Agent Harness for Long-Running Work

CoworkOS keeps agents, people and changing files in the same workflow, making recovery and revision control part of the product.

By Paras Madan | 2026-11-03


CoworkOS is an open-source agent harness described in a November 2 research paper. It coordinates delegated work, preserves task state and lets people edit the same artifacts as agents. The practical question it addresses is how an AI project continues when the work outlasts one conversation.

An agent that produces a first draft has completed only part of a normal project. Someone changes the brief, a long-running process finishes overnight, or a colleague edits the document. Continuing from the right state can matter more than generating another plausible answer.

## How the CoworkOS harness works

The [official repository](https://github.com/RUC-NLPIR/CoworkOS) separates three responsibilities. A main agent chooses the plan and evaluates results. A coordinator manages assignments and execution state. Subagents carry out individual pieces of work.

It also retains job logs and supports shared documents, presentations, CAD models and Blender scenes. Model services are configurable rather than built into a single model release.

This separation is useful because planning and bookkeeping fail differently. A model may choose a poor next step, while a runtime may lose track of whether a step already happened. Improving the prompt addresses only one of those problems. **A persistent agent needs a reliable record of the work, not just a longer conversation.**

## What happens when a person edits the file?

The [CoworkOS paper](https://arxiv.org/html/2611.03178v1) describes revision checks for shared resources. A proposed edit is checked against the version it started from. If another edit has changed that version, the proposal becomes a conflict rather than silently overwriting the newer work.

Imagine an agent preparing a presentation while the founder corrects a revenue number. The important test is whether the next revision preserves that correction, not whether the agent can generate another polished slide. Revision checks provide a concrete place to detect the disagreement before a stale change becomes the final deliverable.

That also suggests a better evaluation for collaborative agents: deliberately interrupt them, change a file and inspect what survives when they resume.

## CoworkOS availability and setup

The repository is MIT-licensed. It documents a desktop installation for Apple Silicon Macs running macOS 14 or later, plus a command-line version for macOS and Linux. Both require Node.js 22.19.0 or newer. Users configure a model endpoint, model ID and API key; the connection test consumes model quota.

Open-source software does not make the underlying model calls free. Teams should account for those calls alongside any external tools and infrastructure they connect. They should also review what data goes to the configured provider before using private project files.

## What the research establishes

The authors evaluate CoworkOS across 12 benchmarks and describe six categories of application work. They also report an automated post-training case study in which an eight-benchmark average rose from 33.19 to 42.95 over five rounds.

Those are preliminary, author-reported results. A training case study does not establish that every installation will improve itself, and a collection of benchmark scores does not establish dependable unattended operation in a company's environment.

For teams evaluating the project, I would start with a bounded task that includes an interruption and a human edit. Measure completion quality, recovery behaviour and model cost together. The useful advance to watch is whether an agent can keep responsibility for a changing piece of work without making the person reconstruct the whole project each time.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI founders turn technical product ideas and launches into clear stories developers can use and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).