Scale AI ships 3,300+ RL environments; some tasks take humans 3 weeks and 500+ tool calls
alexandr_wang · x · 2026-11-06
Scale AI published a major update to its RL environments for agent training, covering 3,300+ tool-use and computer-use tasks already running with customers.
- Enterprise Atlas: built from anonymized real company operating histories — 2,000+ tasks, 4 enterprise environments, 870K artifacts. Example scenarios include a mobile game studio (76+ scenarios, 10 systems, 500+ tool calls per super long-horizon task) plus an accounting firm, mortgage broker, and logistics company. Training Qwen3.8-27B yielded +13pp on Toolathlon and +3pp on AutomationBench.
- Computer-use environments: targeting OSWorld 2.0 and Agent Last Exam, with 1,000 unique tasks averaging 100+ steps each, verified by a rubric-based agent judge, compatible with Linux, Windows, and macOS.
- Environments simulate real systems via MCP servers, CLI, and GUIs; some tasks would take a human team 3 weeks.
More from coding & agent
- A notebook app turns Claude into a tutor on reMarkable and iPad — tokentrillionaire · 2026-11-06
- Measured: in-process BM25 memory costs ~4.6KB RAM per entry, linear growth — Neither-Witness-6010 · 2026-11-06
- Tip: add a phone preview screen to make mobile vibe coding painless — yangyi · 2026-11-06
- Building agents: life-oriented companions beat productivity tools in real use — yangyi · 2026-11-06
- ThinkyMachines' Inkling 2 Small trains models on its own: $426 of compute doubles Qwen score in 4.5 hours — Scobleizer · 2026-11-06
- Skillry ships a Claude Skill recreating Anthropic's homepage hero design — yihui_indie · 2026-11-06