StateAct hits a new OSWorld 2.0 SOTA by using program state instead of pixels
LiJunnan0409 · x · 2026-07-29
StateAct reaches a new OSWorld 2.0 SOTA by prioritizing program state over pixels
The post highlights StateAct, a new approach for long-horizon computer-use agents. Its core idea is to treat program state—files, backends, and DOM—as the primary interface, instead of relying mainly on screenshots.
Key claims from the thread:
- The agent barely looks at the screen, using a GUI specialist only for genuinely visual subgoals.
- That visual fallback is used in just 1.1% of steps.
- On the same Claude Opus 4.8 setup, the method improves OSWorld 2.0 performance, with the post citing gains from 20.6% to 26.9% on binary and 54.8% to 61.x% on partial.
The takeaway is that for long-horizon computer use, the winning signal may be less about better screenshots and more about structuring access to the underlying application state.
More from coding & agent
- Coding agent runs are getting long — one dev suggests breathing, not parallel sessions — teropa · 2026-10-16
- Dev builds MCP layer so Claude Code can work on servers without ever holding SSH keys — NETMESAFE · 2026-10-16
- As coding agents run longer autonomously, one dev's tip: don't parallel-session, just breathe — teropa · 2026-10-16
- 'Refusing to code with AI in 2026 is like refusing Google in 2005' — prasenx · 2026-10-16
- Claude Code adds idle compaction to save 1h prompt cache on idle sessions — EricBuess · 2026-10-16
- Handling agents that crash mid-task: the hard part is the unknown outcome across the network — Responsible-Dare5139 · 2026-10-16