Cowork-Gym: Alibaba Synthesizes 64.9K Agent Work Environments, Boosting Qwen by 8.7 Points
Anni Zou · hf · 2026-10-16
Alibaba's Token Foundry team releases Cowork-Gym, a framework for synthesizing interactive Cowork (collaborative agent) environments at scale—where agents navigate tool/file dependencies in shared workspaces while adapting to changing user requirements.
- Synthesis: A build–execute–revise loop validates and repairs generated environments using execution evidence and review feedback; a stateful user simulator responds via dialogue and direct edits to the shared workspace.
- Scale & training: 64.9K environments across 14 application scenarios, used to post-train 9B/27B/35B-A3B Qwen models with SFT, RL, and on-policy distillation.
- Results: Average gains of 8.7 / 6.4 / 4.1 points across six agentic benchmarks; the 27B ranks second overall and achieves the highest average on PawBench and ClawEval, beating models over 100B parameters.
More from coding & agent
- DIRECT: retrieval-based next-question suggestions for conversational agents — burkov · 2026-10-16
- Claude Code and Codex now autocomplete the user's next action — tokenbender · 2026-10-16
- Hermes Agent ships official Reverse Engineer Anything plugin with one-command setup — Teknium · 2026-10-16
- MLExpert: 35B MoE Agent Trained to Autonomously Handle Machine Learning Engineering — Shangheng Du · 2026-10-16
- MMSearch-OPD: Joint RL and On-Policy Distillation for Long-Horizon Multimodal Search Agents — PekingUniversity · 2026-10-16
- A gallery of 214 projects built on one model shows the 'post your build' ad flywheel — gaganghotra_ · 2026-10-16