Story · arXiv
StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling (arXiv)
paper · Story page
StateM wraps unchanged model weights in an agent-native runtime: durable state, checked transitions and recoverable runbooks. On Terminal-Bench 2.1 it takes GPT-5.6 Sol xhigh to 95.3% raw accuracy across 445 trials, and the frozen profile lifts GPT-5.6 Luna from 76.7 to 85.4%.
In plain words
- Researchers built StateM, software that helps action-taking artificial intelligence systems finish long computer tasks more reliably.
- It saves changing task information so the system does not lose track between steps.
- It separates work into phases and checks that each phase is complete before moving forward.
- It keeps recovery instructions and lessons from earlier attempts available for later runs.
- Tests suggest teams can improve long-task results by organizing execution better, without changing what the underlying artificial intelligence learned.
Appeared in
- Aborted agent branches live on in the KV cache, plus StateM's harness scaling
Aug 19, 2026 · in the sections
Subscribe
Get the brief in your inbox
Pick daily, weekly, or both. Nothing is gated either way: every issue is on the site and in the feeds.
- Weekdays at 8:45am IST, one lead story and 6 to 9 items.
- Sundays, an argued synthesis rather than a recap.
- One click to leave, and quiet days say so in the subject line.