Show HN: Deterministic browser control for AI agents (~90% on Mind2Web) (github.com)

A new tool named Agent Browser Protocol (ABP) has been introduced to enhance deterministic browser control for AI agents, achieving a notable accuracy rate of 90.53% on the Online Mind2Web benchmark. ABP transforms traditional web navigation into a structured, step-by-step process compatible with AI agents, allowing them to interact consistently within a controlled browsing environment. Built on a modified Chromium framework, ABP uses HTTP requests instead of WebSockets, eliminating the need for complex session management. This assures that every action taken by the AI is based on a stable, "settled" page state, complete with screenshots and event logs. The significance of ABP lies in its ability to address the inherent mismatch between asynchronous web browsing and the sequential reasoning of AI agents. By pausing JavaScript execution and incorporating direct event injections, ABP creates a predictable framework for agents, reducing the reliance on heuristic waits and allowing for smoother decision-making processes. This innovative approach not only simplifies the automation of web tasks for AI but also facilitates the generation of fine-tuning datasets for vision-language models. The simplicity of its REST API and the automatic handling of various action states position ABP as a potentially transformative tool for the AI/ML community, bridging the gap between web automation and intelligent decision-making.