Browser Testing Agent Loop

Traditional E2E test suites only catch regressions in scenarios a human has already thought of. An agent with a vision model catches the messy class of bugs nobody scripted: broken layouts, accessible-but-invisible elements, misleading copy. The report-then-continue pattern prevents one bug from blocking discovery of others, and the screenshot-plus-DOM snapshot gives developers a concrete repro artifact, not a hallucinated bug description.

1) Spin up a headless browser (Playwright). 2) Navigate to the start URL; capture full-page screenshot + accessibility tree. 3) Feed both to the vision LLM with a system prompt: 'you're a QA engineer; choose the next action that explores a new area of this app'. 4) Parse returned action: click / type / navigate. 5) Execute; re-capture. If a condition from the user story is violated (404, network error, visual regression vs a baseline), emit a bug record with: title, reproduction steps from the action log, severity, attached screenshot. 6) Continue until the budget is hit or every link has been touched. 7) Output: bugs.jsonl + a browsable HTML report.