Browserbase: browser agents that write code beat pixel-clicking CUA models
adnan_hashmi · x · 2026-09-10
Browserbase's engineering team reflects on computer-use models (CUA) since OpenAI's Operator launch in January 2025, and argues for a key shift: letting agents write code to drive the browser instead of relying purely on vision-based coordinate clicks.
- CUA pain points: models trained on fixed viewports (e.g. 1280p) go effectively blind when resized and misclick; vision-only approaches can't reach components invisible in screenshots, which modern web apps are full of.
- Their finding: getting out of the way and letting models do what they're good at—writing code—improves performance across all domains.
- The post also recaps their Astra architecture, earlier open-source implementation, and notes they helped Anthropic and Google DeepMind train and evaluate computer use in Sonnet/Opus and Gemini 2.5 Pro.
More from coding & agent
- NVIDIA joins Rust Foundation as Platinum member, pushes official Rust support across CUDA — vinodg · 2026-09-10
- Notion CEO-shared take: own your context, rent the intelligence — ivanhzhao · 2026-09-10
- Voice in the Codex app is great for math study, but bugs pile up in heavy use — SIGKITTEN · 2026-09-10
- Quantum sensor team uses GPT-5.6 Pro + Codex on Inverse Galois Problem, ranks 14th on IGP24 leaderboard — paulfinneyx · 2026-09-10
- Dev Recreates the MacBook Folding Animation Entirely in the Browser — jh3yy · 2026-09-10
- AI gateway vs MCP gateway: do production agent stacks actually need both layers? — Purple_Morning_8735 · 2026-09-10