QA Engineer

Problem

Small teams ship without end-to-end tests because writing and maintaining them is slow.

Try it live

Switch on bugs in the staging shop, such as a coupon that applies twice, a checkout that fails in one browser, missing alt text or a cart count that is off by one. Give a goal like buying two bags with WELCOME10. Watch the run step by step, then read the bug report, look at the screenshots and read the generated test.

Proves

Browser agents, test automation, and sandboxing done properly.

Signal chain

1. plan from goal
2. run plan in Playwright
3. re-plan on failure
4. catch errors + axe checks
5. bug report
6. generate test
7. run red, then green

Stack

Node · TypeScript · Playwright · axe-core · BullMQ · a locked-down sandbox container

Engineering highlights

• A generated test is kept only if it fails with the bug switched on and passes with it off.
• A container for every run would need the Docker socket, which the platform never exposes, so each run gets a fresh browser context in one locked-down sandbox container, closed at three minutes and restarted every few runs, where the browser can reach only the staging shop: that closes the server-side request forgery hole “test any URL” demos leave open.
• It plans the whole test in one call and asks the model again only when a step fails, which keeps a run to seven calls at most and makes it repeatable.