A general-purpose long-task agent shell from teams including Renmin University's NLPIR, whose core idea is to separate "strategy" from "execution management": the model decides what to do, whom to delegate to, and whether to accept the results, while the runtime handles state persistence and feedback. The three-layer architecture consists of a main agent (planning, delegation, evaluation), a coordination runtime (dispatching subtasks, tracking lifecycles), and optional sub-agents (executing scoped tasks), supporting both synchronous and asynchronous modes. The shared application layer integrates CLI, search, files, MCP, document editors, presentations, CAD, and Blender, allowing humans to participate via conversation or by directly editing the same document or scene; hierarchical workspaces and persistent semantic memory enable files, key decisions, and long-running job logs to resume after interruptions.
Likes 88 · Code Stars 78
Reasons
- The authors claim it performs comparably to frontier agent applications like Codex and DeepSeek Harness across 12 benchmarks and 6 categories of real-world applications.
- Long-term automation case study: Using GLM-5.3 as the backbone, CoworkOS performed 5 rounds of automated post-training on Qwen3.6-35B-A3B (analyzing failures, modifying data, training, validating). The authors report that the average score across 8 benchmarks rose from 33.19 to 42.95 (+9.76).
- Easy installation: MIT license, npm package coworkos-app, with a macOS desktop version and macOS/Linux command-line versions. Model services are configured by the user, and workspaces and execution records remain local. It ranked fourth in likes on the day of release (indicating attention level only).
Usage
- Desktop version (Apple Silicon, macOS 14 or later, Node.js ≥22.19.0): npx coworkos-app@latest desktop install.
- Command-line version: npm install -g coworkos-app. After entering the project directory, run coworkos init to configure the model service address, model ID, and API Key. Run coworkos doctor --test-connection to check connectivity (this consumes one quota unit), then run coworkos to describe the task.
- After exiting, run coworkos --continue in the same directory to resume the previous session. See the project page showcase for an online demo.
Risks
- It is only a shell and does not include models; performance depends on the model service and quota you connect to. Comparisons with Codex and scores from the post-training case study are reported by the authors.
- The desktop version currently supports only macOS on Apple Silicon. Agents can execute commands and modify files. The README notes that third-party components have separate license exceptions, which users must verify themselves.
- The automated post-training case study requires training compute resources, and the README does not provide reproduction scripts.