CMU's Gym-Anything Turns Any Software Into an Agent Environment: 200+ Apps, 10,000 Tasks, Open Source
wellecks · x · 2026-09-30
CMU's NeurIPS-accepted paper "Gym-Anything: Turn Any Software into an Agent Environment" automatically converts any desktop software into a computer-use agent training/testing environment, fully open source.
Key facts
- Scale: 200+ real-world apps (GIMP, Blender, NinjaTrader, OpenEMR...), 50 tasks each, 10,000+ tasks/environments; covers all 22 SOC occupation groups.
- Cross-platform: Linux (desktop), Windows (via QEMU), Android (emulators), unified API, all in sandboxed VMs with realistic data.
- Core insight: setting up computer-use environments is itself a coding/computer-use task, so the whole pipeline can be automated.
Pipeline
- SELECT: decompose US GDP data into per-software economic contribution to pick 200 high-impact apps.
- CREATE: a coding agent installs, configures and populates each app; an independent audit agent verifies against quality checklists until production-ready.
- SCALE: seed tasks are created by an agent actually operating the software, amplified via LLM; each task gets a checklist-based verifier with privileged information.
Also releases the harder CUA-World-Long benchmark, plus paper, code, interactive docs and a leaderboard.
More from coding & agent
- Denying GPT-6.1 Sol write access as orchestrator cut coding costs 77%, at 6x runtime — GapNew4766 · 2026-09-30
- 8 research agents self-train a 30B model for 144 hours in RSIArena livestream experiment — my_cat_can_code · 2026-09-30
- OpenAI's Nan Yu: a boss agent running other agents is just one agent with extra steps — victor_explore · 2026-09-30
- Open-source MCP server lets agents query 124.6B TikTok data points — operatorarkay · 2026-09-30
- Dev argues truly always-on autonomous agents have never actually been tried — jacob_posel · 2026-09-30
- Skill scaffold template: same shape, faster review, fewer broken CIs — blaizedsouza · 2026-09-30