LLM Multi-Agent QA System
Four agents (planner, executor, verifier, supervisor) driving Android UI flows to deterministic completion.
- Independent build
- Independent
- 2025
- Shipped
Problem
Most agent tests fail in boring ways: a bad click, a missing state check, or a recovery path that was never designed.
Approach
I built a multi-agent QA system for Android flows. One agent plans, one executes, one checks state, and one handles recovery when the app does something unexpected. Splitting those roles mattered more than making any single one smarter.
What I built
An agentic ADB pipeline that converts natural language into Android actions using episodic memory, with a planner, executor, verifier, and supervisor coordinating over a message bus.
Architecture
Planner agent decomposes natural-language test goals into executable UI steps, replanning when the app diverges.
Executor drives Android Debug Bridge with GPT-4o and OpenAI Vision reading the screen, while the verifier checks state transitions.
Supervisor logs decisions and owns retries, failure recovery, and deterministic completion criteria.
Impact
Achieved more than 99% deterministic execution across complex UI flows in internal benchmarks.
>99%