XNews

Browser Use and the Bitter Lesson: Simplifying Harness for Browser Agents

Browser Use and the Bitter Lesson: Simplifying Harness for Browser Agents

Browser Use co-founder Greg PR shared lessons from two years of evolution: as models get stronger, the harness should be stripped down. The team moved from predefined actions to letting agents write JavaScript, then to giving them raw CDP access. This reduced token consumption by 60% for Opus 4.8 and improved reliability, following Sutton's Bitter Lesson in the browser agent domain.

Key insights

Evolution from Predefined Actions to Raw CDP

Browser Use started with human-defined state and action spaces (click, input, scroll). In 2025, they introduced JavaScript execution and persistent notebooks, allowing agents to write programs. Later, they removed predefined state extraction, letting agents directly access CDP to decide what to observe. The final step was to reuse existing coding agent harnesses (Pi, Codex, OpenCode) instead of building custom loops.

Key Evidence: Token Reduction and Reliability Gains

Replacing 12 browser tools with a single browser_exec tool reduced token consumption by 60% for Opus 4.8 and 66% for Kimi K3, while both completed all 18/18 task runs. This demonstrates that giving agents more freedom (writing code) improves both efficiency and reliability simultaneously. The team also found that CDP can penetrate closed shadow roots where Playwright's locators fail.

Three Principles for Agent Harness Design

The article distills three transferable principles: reuse proven agent harnesses; expose the simplest possible low-level interface that models can use well; let models choose their own observation and actions. The author states this applies to all agent domains (browser, computer, text, music). This aligns with the 2026 industry consensus of minimizing harness engineering.

Source posts