Browser Agents' 'Bitter Lesson' from Browser Use

Browser Use co-founder Greg shows that as models improve, the harness should be stripped down to the raw CDP interface. Over two years, they moved from predefined actions to letting models write code and directly access the browser's native protocol, improving efficiency and reliability.
Key insights
Evolution of Harness Design
Initially, the harness defined state and action spaces (click, input, scroll) because GPT-4o was not trained for computer use. Then they introduced JavaScript execution, replacing 12 browser tools with one browser_exec tool, reducing token consumption by 60-66% while completing all tasks. Finally, they removed predefined state observation, letting models connect directly to CDP.
Why CDP Over Playwright
CDP is Chrome's native protocol, below Playwright's abstraction. Playwright's locators cannot penetrate closed shadow roots, while CDP can. The team also stopped building their own agent loop, reusing proven coding agents like Pi, Codex, and OpenCode instead.
Three Principles for Agent Harness
1) Reuse proven agent harnesses. 2) Expose the simplest low-level interface the model can use well. 3) Let the model choose its own observation and actions. This principle applies beyond browsers to computer, text, music, and other agent domains.