LinkedIn built an autonomous QA Agent that uses Vision-Language Models (VLMs) and a multi-model architecture to test iOS, Android, and Web apps without relying on brittle selectors. The system uses a hybrid 'System 1/System 2' approach: deterministic replay for known UI paths and VLM-based planning when the UI changes. Multiple evaluators act as guardrails to catch hallucinations and false positives. A golden dataset framework enables reproducible, offline evaluation and safe experimentation with new model configurations. The agent has found over 200 bugs, caught critical regressions in revenue flows, and enabled non-engineers to author tests via natural language. Future plans include shifting testing left into the development cycle, automated triage, specialized sub-agents for accessibility and i18n, and exploratory testing.
