Salesforce’s StateAct lifts Opus 4.8 on OSWorld 2.0 with state-first agents
Salesforce · hf · 2026-07-28
Salesforce proposes StateAct, a code-first harness for long-horizon computer-use agents that works on program state before pixels.
- The paper argues screenshots are lossy: different underlying states can look identical, while code can directly inspect and modify task data in files, backends, and DOM.
- StateAct splits the system into a main agent that manipulates program state with code and a GUI subagent that only handles the few click-based subgoals.
- On OSWorld 2.0, it boosts Claude Opus 4.8 from 20.6% to 26.9% binary success and from 54.8% to 61.6% partial success, while cutting task cost by about 9× versus screenshot-only driving.
- A code-only variant without the GUI subagent reaches only 45.9% partial success, showing the GUI branch still matters.
- The authors frame this as state-grounding: grounding action, verification, and memory in program state shifts the bottleneck from perception to reasoning.
More from coding & agent
- Mnemos is getting a rebuilt desktop app, new MCP tools, and a redesigned platform — RileyRalmuto · 2026-07-28
- Perplexity Computer and Comet Assistant built 7 GA4 reports in 90 minutes — gaganghotra_ · 2026-07-28
- Codex hits a session limit after one shell command in a coding workflow — AIandDesign · 2026-07-28
- Distributed Systems open-sources DSCO, a C-written agentic LLM CLI with MCP and swarms — arthurcolle · 2026-07-28
- Local coding-agent roundup says Gemma 4 26B-A4B beats dense 24B models in refactoring tests — craigsdennis · 2026-07-28
- Credit assignment, not better data or entropy, looks like the bottleneck for agentic RL — burny_tech · 2026-07-28