agent-qa
A self-improving QA test harness that runs natural-language test scripts for web and mobile apps, re-observing the interface to recover from failed steps and building memory across runs to catch regressions.
agent-qa is an agentic QA harness for testing web and mobile applications. Tests are written as natural-language actions and assertions, and the agent works from visible roles, labels, and screen state rather than brittle selectors. When a step such as a click or a fill fails, it re-observes the interface and tries a different path within the same run, so tests recover from UI drift and flaky interactions.
Across runs it builds execution memory from product, suite, and test observations, and curates memory from steps that needed healing so future runs avoid the same mistake. An action cache reuses validated plans to cut planner work and token usage. Hooks run in sandboxed Docker containers to set up environments, seed fixtures, and tear down state.
It ships a dashboard and a command-line interface, plus MCP and skills for coding agents, and works with OpenAI- and Anthropic-compatible model endpoints, Gemini, local models, and Codex or Claude Code subscriptions. It is written in TypeScript.