Goal-driven mobile app test runs
Give an agent a goal like 'request a ride and pay' and let it drive the app screen by screen, reporting where it gets stuck.
The use case
Scripted mobile UI tests break every time a screen layout changes, and maintaining them across many markets is expensive. Uber's DragonCrawl uses a language model to read each screen and decide the next tap toward a goal, so the same test runs across cities and languages without per-market scripts. When the agent cannot finish, the run is flagged for a person. QA engineers still triage failures, since a stuck run might be a real bug, a confusing screen, or the agent's own mistake, and they keep a smaller set of scripted tests for critical paths.
Example workflow
- Write plain-language goals for key user journeys
- Run the agent on emulators or devices against a test build
- Agent reads each screen and chooses the next action
- Collect screenshots and logs for failed runs
- QA triages failures into bugs or agent errors
Tools
- LLM API
- Android emulator
- Mobile test harness
Risks and guardrails
- Agent may take odd paths that real users never would
- Never point it at production accounts that can spend money
- Flaky agent results need human triage
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- Uber engineering blog: DragonCrawl mobile testing: www.uber.com/blog/generative-ai-for-high-quality-mobile-testing