1. 01Agentic QA

   Write the test the way you would explain it

   Numbered steps in plain words, with a frontmatter for the platform and the state it starts from. Verify and extract are steps too; values you extract become variables for later steps, and a state you save can start the next test. The same file is understood by the editor, the worker and the repository.

2. 02Agentic QA

   A browser agent runs it, in Chrome, on every push

   Six cooperating agents drive a real browser over its debugging protocol: one plans, one acts, one asserts, one judges the interface, one extracts, one plays the persona. Nineteen actions, including throttling the network and the CPU to see what a customer on a bad connection sees. No selectors to maintain: a redesign does not break the test, a broken flow does.

3. 03Agentic QA

   Evidence, streamed live

   A session replay, the console and the network log stream while the run happens; a late joiner catches up, then follows live. When a step cannot be verified, the run says so with a reason instead of faking a pass. The probe engine handles plain HTTP checks without a model or a browser.

4. 04Agentic QA

   Evidence a reviewer would ask for

   Open any step and see exactly where the agent clicked, drawn on the page it clicked, with a crop of the element beside each action. A design reviewer looks at every screen the run visits at the bar of “would Apple ship this exactly as it is” and lists what it would hold back — misrendered or misaligned elements, confusing flows, sloppy copy — numbered on the page. Every request opens to its headers, body and timing; every step keeps its console, its document and the browser’s load. Findings are warnings: they never fail a run.

5. 05Agentic QA

   The agent’s loop: the push that came from an agent gets the same run

   A coding agent’s push triggers the same QA run as a person’s. The run, its replay and its verdict sit in the hub on the branch, where the reviewer of the agent’s change request sees them before approving. Runs also start by hand or on a schedule; a project can pause new runs without cancelling queued ones. On Android a device agent drives an emulator reset to its clean snapshot before every test — the keyboard, permission prompts, rotation — with the app’s log, its crashes and its load beside each step.

Coming soon: the device agent on iPhone and iPad simulators · automatic scheduling for test groups

The full reference lives in the docs.