autobrowse

AutoBrowse is a self-improving browser automation skill that uses an iterative research loop to build reliable web navigation strategies. It works by running a browsing task, analyzing what happened, and automatically refining the navigation instructions until the task passes consistently. The skill separates concerns between an inner agent that executes the browsing and an outer agent that learns from the results, creating a feedback loop that improves automation reliability over time.

The skill supports both local and remote browser environments, making it adaptable for bot-protected sites that require sophisticated browser fingerprinting. It can run single tasks sequentially or process multiple tasks in parallel using sub-agents, with configurable iteration counts to balance training time against reliability. An optional browser-trace integration provides deep CDP-level diagnostics including network traffic, console logs, and page lifecycle events for debugging complex navigation issues.

AutoBrowse is designed for developers and automation engineers who need to create maintainable browser automation workflows without manually debugging brittle selectors or navigation logic. It's particularly valuable for testing web applications, scraping data from dynamic sites, automating repetitive web tasks, and building robust end-to-end testing strategies. The skill stores all training artifacts in a local workspace directory, keeping experimentation organized and separate from system files, with a graduation step that exports successful skills for reuse.