A developer has forked Chromium to create agent-browser-protocol (ABP), a system that synchronizes browser state with AI agents at every step. The core problem ABP solves is that LLMs often fail when interacting with web pages because they reason from stale browser states.

Traditional browser automation forces AI agents to race against a live, continuously updating page. ABP eliminates this by freezing JavaScript execution and rendering after each action, then capturing the resulting state with screenshots and event logs. The approach transforms web navigation into a discrete, multimodal chat loop that better matches how LLMs already work.

ABP achieves 90.53% accuracy on the Online Mind2Web benchmark, demonstrating that modern LLMs understand websites but need better tools to interact with them. The system provides a simple REST API that returns settled page states, screenshots with cursor positions, and structured event logs for navigation, dialogs, file choosers, and downloads. By embedding an HTTP server directly in the browser process and controlling execution at the engine level, ABP gives AI agents deterministic step boundaries and prevents the race conditions that plague current browser automation.