Getting started
Autonomous agent workflow
Pair a UI driver with Genie so the agent can reproduce, fix, and verify the work.
Always use Genie with a UI driver.
| Step | Tool | Result |
|---|---|---|
| Open the app and perform the flow | agent-browser or agent-device | The real user action runs. |
| Read what happened underneath | Genie | The agent gets components, causes, cache, routes, effects, and timings. |
| Change the source | Coding agent | The smallest evidence-backed fix is made. |
| Repeat the same flow | agent-browser or agent-device | Visible behavior, layout, focus, and accessibility are checked. |
| Compare live evidence | Genie | The agent checks the same selected runtime evidence again. |
| Restore test state | Genie and the UI driver | The developer gets a clean app. |
Web
agent-browser drives and checks the page
↓
Genie explains React, Query, Router, memory, and frames
↓
the agent edits, repeats, and verifiesRun the full browser loop in First live check.
React Native
Use agent-device for taps, typing, screenshots, and native accessibility. Use Genie for React, Query, Router, hooks, effects, and renders.
Run the full device loop in Verify a native app end to end.
Done means both pass
The work is complete only when:
- The UI driver confirms the feature still works.
- Genie supports the result with the selected evidence.
- Genie reports enough coverage for that claim. Incomplete or truncated coverage needs another run.
- Temporary Query state and React overrides are restored.
Genie cannot prove a path it did not observe. Match each claim to a visible assertion, a focused Genie read, and that read's coverage fields.
Return the result
Never end after the plan or tool calls. Return a short report:
- What the agent reproduced.
- What the evidence proved, including coverage limits.
- What changed.
- Which live checks and tests passed.
- What the developer should do next, or
nothingwhen the work is complete.
If no change was made, say why. If verification failed, show the failed check and the next action.