agent-browser

Browser automation for AI agents is achieved by driving a CLI that leverages accessibility tree snapshots and ref-based element selection, ensuring stable interactions. This approach enables deterministic navigation and multi-step workflows across dynamic web applications.

Parallel browser automation sessions are fully supported through session isolation, allowing you to run multiple independent browser contexts simultaneously. This prevents cross-contamination and supports multi-session testing across different workflows.

Ref-based element selection in web automation works by capturing accessibility tree snapshots of the page, then using the assigned refs to interact with specific elements. This method provides deterministic targeting that remains stable across complex single-page application DOM changes.

The best way to capture page state for AI reasoning is using snapshot-based state capture, which records the accessibility tree of the current browser session. These snapshots support AI decision-making, debugging, and data extraction during autonomous workflows.

Browser automation using accessibility snapshots and ref-based selectors works reliably with dynamic single-page applications. The snapshot approach decouples interactions from volatile DOM structures, ensuring stable multi-step navigation and form completion.

Accessibility snapshots are used instead of standard DOM selectors to provide deterministic, stable element selection for AI agents. This method reduces fragility from dynamic DOM changes, ensuring reliable repeatable browser automation and session isolation.