CUAWright: Terminal-Only Computer-Use Agent Beats GUI Harnesses, Cuts Cost 37.5%
ysu_nlp · x · 2026-10-07
- The CUAWright paper argues computer-use agents shouldn't be locked into fixed GUI harnesses; instead, give them just a terminal (3K lines of code) with runcommand as the sole action interface.
- A file system serves as the evolvable space where the agent organically builds its own tools and manages context.
- Results: +33.2% relative partial reward on OSWorld 2.0 while cutting estimated cost 37.5% vs the published GPT-5.5 baseline; beats GUI-native harnesses by 4.7% on Online-Mind2Web and 44.0% on the long-horizon Odysseys benchmark.
- Gains extend to CAD tasks: 8.1%-41.6% relative improvements on CADGenBench and BenchCAD.
More from coding & agent
- Dev Recreates All Adobe Apps with Opus 5.5, Ports Them to Rust, Open Sources — SumitGup · 2026-10-07
- AgentID Launches: Email-Based Identity So AI Agents Can Sign In as Themselves — omarsar0 · 2026-10-07
- agentid launches as 'sign in with Google' for AI agents, with email and logins — prasenx · 2026-10-07
- AI Engineer talk: LLM judge must match humans 80% before anything ships — TejasKumar_ · 2026-10-07
- Multi-agent pipelines break when two correct agents disagree on what 'done' means — ClickOk5811 · 2026-10-07
- MCP server for finding paid x402 APIs: 7,400 listings, 1,877 verified with real USDC test calls — SashSail · 2026-10-07