The bitter lesson of browser agents: as models improve, strip away the scaffolding

The bitter lesson of browser agents

Browser Use CTO Gregor Zunic walks through three rewrites of their agent architecture over two years. They started by feeding GPT-4o a predefined page state and a fixed action menu. By September 2025 they let the model write JavaScript directly, cutting token usage by over 60%. The next bottleneck was the observation layer—their state extraction missed cookie buttons and shadow-DOM dropdowns. Now they hand raw CDP access to the model so it sees the page and writes its own execution code, keeping only a thin harness. The post does not disclose benchmark numbers for the current architecture.

Why it matters: First-hand architecture postmortem from Browser Use's CTO — three rewrites and a 60% token cut give it substance. It's an engineering experience share, not a product launch, so it doesn't hit the 85+ band.