Browser Use Releases 'Browser Harness' Allowing Direct CDP Access by Applying the 'Bitter Lesson'

The Browser Use team has released 'Browser Harness', a new architecture that applies Sutton's 'Bitter Lesson' by discarding the previous approach of relying on predefined states and action spaces, and instead allowing direct access to the Chrome DevTools Protocol (CDP).

Architectural Changes and Rationale

The previous approach limited the information visible to the model to accessibility trees, etc., which caused bottlenecks in recognizing complex DOM structures such as closed shadow roots or iframes from other origins. The new approach grants the model direct access to CDP, allowing it to write and select the necessary observations, such as DOM inspection and screenshot capture, directly in code. While Playwright has limitations, such as locators being unable to see closed shadow roots, CDP allows for direct inspection of these internal structures.

Tool Integration and Performance Improvements

The Hermes agent replaced the previous 12 browser tools with a single browser_exec. Stability was ensured by reusing proven coding agent loops from Pi, Codex, and OpenCode. Performance test results showed that when using the Claude Opus 4.8 model, token usage decreased by 60%, and for the Kimi K3 model, it decreased by 66%. Both models successfully completed all 18 runs (3 runs each for 6 tasks).