Paper review · November 4, 2026
CoworkOS: an agent coworker harness for real-world work automation
CoworkOS:Scaling Agent Coworkers for Real-world Automation
A three-layer harness that leaves strategic decisions to the model and execution management to the runtime supported long-running work, and automated post-training with it raised Qwen3.6-35B-A3B's average score across 8 benchmarks from 33.19 to 42.95.
Problem
Productivity work that produces software, documents and designs needs two things at once: the ability to complete complex outputs at high quality, and the ability to monitor training runs or projects lasting several days and keep making judgments. Harnesses that support both in a form where people can step in are still lacking.
Method
CoworkOS has a three-layer structure. The Main Agent handles overall planning and final review of outputs. The Coordinator Runtime manages subtask assignment, lifecycle tracking and result delivery as persistent state. Subagents carry out delegated tasks in isolated contexts. Delegation can be synchronous or asynchronous, and asynchronous results accumulate in the Main Agent's Inbox and are taken into account at the next decision point. In the shared application layer, such as document editors, PPTist, CAD and Blender, people and agents can directly view and edit the same outputs. A hierarchical workspace, persistent semantic memory and Jobs that record state and logs make it possible to resume interrupted work.
Results
- The authors reported performance competitive with major agent harnesses such as Codex, DeepSeek Harness and OpenHands on 12 benchmarks and real-world applications in 6 categories (web frontend, games, office presentations, interactive visualization, CAD and Blender rendering).
- CoworkOS, using GLM-5.3 as the foundation model, ran five rounds of automated post-training on Qwen3.6-35B-A3B. The average across 8 benchmarks rose by 9.76 points, from 33.19 to 42.95.
- In this RSI (recursive self-improvement) case, CoworkOS automatically and repeatedly carried out weakness analysis, data pipeline optimization, data synthesis and verification, and training monitoring and log management.
Limitations
Within the provided text, detailed scores and comparison conditions for the 12 benchmarks could not be found. The automated post-training case covers only five rounds with one foundation model (GLM-5.3) and one target model (Qwen3.6-35B-A3B). It is also not confirmed within this text whether the effect of human involvement was evaluated quantitatively.
Why it matters
For organizations that want to go beyond coding agents and hand off documents, 3D and long-term monitoring, it shows a practical design. Examples include splitting context and delegating to Subagents, persistent state and Job logs, and a collaboration style in which people directly edit outputs. It also shows a case in which an agent automates model training itself.
Key terms
- Agent harness
- A runtime around the model that handles tool calls, state and context management, and execution feedback, turning the model's reasoning into actual actions.
- Coordinator Runtime
- A control layer between the Main Agent and Subagents that manages task assignment, execution state and result delivery as persistent state. It lets work continue after an interruption.
- RSI(Recursive Self-Improvement)
- A process in which a system repeatedly analyzes its weaknesses, carries out improvements, and feeds the results into the next round of improvement.
Original
Authors: Jiajie Jin, Guanting Dong, Chengjun Pan, Xin Deng, Zhuohui Sheng, Guoqiang Wang, Tao Gui, Yutao Zhu, et al. · Posted 2026-11-02 · License CC BY-NC-SA 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.