Browser Use deletes its own abstractions and lets the model drive Chrome directly

Zunic traces four stages his team moved through since launching Browser Use in November 2024. First, a fixed action space of click, type, and scroll, with the harness deciding what the model could see. Second, replacing those fixed actions with model-written code, starting with JavaScript execution in September 2025 and persistent notebooks that October. Third, removing the harness's predefined view of the page entirely and giving the model direct access to the Chrome DevTools Protocol, so it chooses whether to take a screenshot, query the DOM, or inspect a single frame, rather than working from a list of clickable elements the harness decided to show it. Fourth, dropping their own agent loop and running that CDP access inside existing coding-agent harnesses such as Pi, Codex, and OpenCode instead.

For the second stage, Zunic points to Hermes, an agent built by Nous Research that switched from twelve separate browser tools to one browser_exec call. In a controlled comparison of six tasks run three times each, mean token use fell 60 percent on Opus 4.8 and 66 percent on Kimi K3, and both models still solved all 18 of 18 runs. Browser Use's case study page reports the same pattern held at production scale: task tokens fell 48 to 66 percent across 204 live runs after the switch. Zunic's summary of the underlying pattern: "As models change, rethink the harness."

Why it matters: Browser Use measured lower token use after removing a fixed tool list and a predefined view of the page. The controlled comparison retained an 18-for-18 solve rate, and the production case study covered 204 live runs. Teams using similar browser-agent abstractions can now compare those results with their own harnesses.