Autonomous QA Agent Suite

Quality engineering absorbs enormous effort in three places that depend on judgement — triaging inbound defects, working out why a test failed, and exploring an application for the defects nobody wrote a test for. That dependence on judgement is exactly why they stayed manual.

• Bug Triage Agent: ingests defect exports, normalises them, classifies severity P1–P4 against an enterprise LLM endpoint, and produces a structured triage report — orchestrated with LangGraph.js, persisted in PostgreSQL, traced in LangSmith so every decision can be replayed.
• RCA Agent: parses pytest output, JUnit XML, stack traces, and logs; correlates failures against Git history; retrieves similar past failures from a vector index; tracks flakiness over time. A confidence gate decides whether the first answer is good enough — below threshold, a ReAct agent picks up tools and works the problem further.
• Exploratory Testing Agent: navigates a web app like a senior QA engineer and detects four defect classes at once — broken flows via vision analysis of screenshots, JavaScript errors via Playwright listeners, WCAG violations via axe-core in the live DOM, and visual regressions via pixel-diff against baselines.

The three most judgement-heavy activities in the QA lifecycle now have an autonomous first pass. Findings arrive in the formats teams already use — triage reports, RCA narratives with cited evidence, colour-coded Excel defect logs — not raw model output.

Four distinct defect classes detected in a single exploratory run