StateAct: Program State Cuts Agent Costs by 9x, Sets New SOTA
_akhaliq · x · 2026-07-28
While current computer-use agents rely heavily on vision models to read screen pixels, StateAct proposes a counterintuitive approach: using program state (such as files, backends, and DOM) as the primary interface, only invoking a GUI specialist for the 1.1% of steps that truly require visual judgment.
This "look less" approach sets a new SOTA on the OSWorld 2.0 benchmark. Using the exact same Claude Opus 4.8, the method increased task success rates from 20.6% to 26.9% while drastically reducing token consumption (from 224K to 100K), slashing cost-per-task from $72 to $7.8.
Furthermore, the framework is model-agnostic; dropping in an internal 31B model (SFR-CUA) yielded significant leaps across multiple benchmarks.
More from coding & agent
- Vibe coding looks unserious until people start shipping real apps with it — alexmacgregor__ · 2026-07-28
- Open-source AI Skill turns The Art of War into a structured decision workflow for Claude Code — yangyi · 2026-07-28
- Hugging Face open sources a real-time speech-to-speech voice-agent pipeline — anselm · 2026-07-28
- Who is liable when an AI agent uses non-commercial software on your behalf? — gabepgomes · 2026-07-28
- Voice may replace keyboards as the main interface for managing agents and computers — noahsolomon · 2026-07-28
- Custom scripts often beat waiting for a better MCP, says one AI workflow user — DavidWells · 2026-07-28