XNews

Browser agents: harness engineering should be minimized

Browser agents: harness engineering should be minimized

Browser Use co-founder Greg PR argued that as models become more capable, harness engineering around browser agents should be stripped down. Over two years, the team moved from predefined actions to giving models direct access to Chrome's CDP, reducing token consumption by 60% while maintaining reliability.

Key insights

Evolution from predefined actions to raw CDP access

Browser Use started with human-defined state and action spaces (click, type, scroll). In 2025, they introduced JavaScript execution and persistent notebooks, allowing models to write programs instead of picking from a menu. This reduced token consumption by 60% for Opus 4.8 and 66% for Kimi K3 while completing all 18/18 tasks. Finally, they gave models direct CDP access, removing predefined state extraction.

Three principles for browser agent harness design

The author distilled three principles: reuse proven agent harnesses (Pi, Codex, OpenCode), expose the simplest possible low-level interface (CDP over Playwright), and let the model choose its own observation and actions. This is an application of Sutton's Bitter Lesson: human-designed intermediate representations become liabilities as models improve.

Source posts