CoworkOS:Scaling Agent Coworkers for Real-world Automation

CoworkOS. Scaling Agent Coworkers for Real-world Automation. By Jiajie Jin and thirty-nine others. Published on the second of November.

Cowork agents are moving human and AI collaboration toward real-world automation, where the output is an artifact: software, a design, something delivered. That kind of productivity work asks for two things at once. The delivery on a complex task has to be high quality, and someone has to keep monitoring and making decisions over a long horizon. Doing both together is still an open challenge. Strong delivery on one task is only half of what a coworker has to do.

CoworkOS is a general-purpose harness for long-horizon work, and its central move is a separation of duties. Task strategy belongs to the model. Execution management belongs to the runtime. The model decides what action to take, whether to delegate, and whether to accept a result. The runtime keeps the state and carries the feedback. Each side holds a stable responsibility, rather than one component trying to do everything.

That separation is built as a three-layer architecture. Work can be executed directly, or handed off to Subagents, and the handoff can run synchronously or asynchronously. Above that sits a shared application layer. Humans and agents look at and revise the same evolving documents, presentations, CAD models and Blender scenes. So a person can join in two ways: by talking to the agent, or by editing the work itself.

Continuity is the other half. To keep work going across agents and across sessions, CoworkOS adds hierarchical workspaces and persistent semantic memory. Working files, key decisions and long-running Job logs stay shareable and reusable, and they can be resumed after an interruption instead of rebuilt from scratch.

The evaluation spans twelve benchmarks and six categories of real-world applications. Across them, CoworkOS is reported as highly competitive with frontier agent applications, with Codex and DeepSeek Harness named as examples.

There is also a sustained-work case study. With GLM five point three as its backbone, CoworkOS ran RSI-driven automated post-training on Qwen three point six thirty-five B A three B. That ran for five rounds. Averaged over eight benchmarks, the trained model's score went from thirty-three point one nine to forty-two point nine five, a gain of nine point seven six points. Here the harness is not just answering a task. It is running a multi-round training project on another model.

The authors present CoworkOS as a path toward general cowork-agent intelligence. The design bet underneath is a division of labour. Let the model own judgment, let the runtime own memory and bookkeeping, and give people a way into the same files the agents are changing. Whether that holds up is exactly what the long-horizon case study is meant to show.