OpenCUA author: the 450-step Paint demo we rejected 2 years ago is now doable by agents
andersonbcdefg · x · 2026-09-06
OpenCUA co-author Bowen Wang recalls rejecting a human demo submitted two years ago that spent 300 steps drawing a sunflower in Windows Paint and 150 more coloring it — deemed an insane task useless for training computer-use agents.
Two years later, he's watching Astra take a portrait photo and autonomously draw the person inside Apple Notes on a Mac. "The game has changed."
OpenCUA itself is an open-source CUA framework from HKU's XLANG Lab, Moonshot AI, Stanford and others, featuring an annotation infrastructure, the AgentNet dataset spanning 3 operating systems and 200+ apps/websites, and a pipeline converting demonstrations into state-action pairs with reflective long chain-of-thought reasoning.
More from coding & agent
- Codex launches Astra: notes persist across context windows without lossy summaries — CShorten30 · 2026-09-06
- yacineMTB explains his AI CAD stack: dingcad, a JavaScript manifold wrapper — PeterDiamandis · 2026-09-06
- AI agent weaponizes live Chrome 0-day exploits autonomously in under 24 hours — cyb3rops · 2026-09-06
- Coinbase Payments MCP gives Claude agents a self-custody wallet with spending limits — kleffew94 · 2026-09-06
- Danny Postma releases free skill to fix AI's terrible landing page designs — rachel_l_woods · 2026-09-06
- yacineMTB demos voice-driven agent workflow over Tailscale and tmux — yacineMTB · 2026-09-06