Notes from the field on test automation, agentic QA, and shipping quality software.
AI reviewers go first on every PR, flag mechanical issues before a human opens the diff, and cut review time roughly in half. Here is how.
A RAG answer can read perfectly and still be wrong. How to test retrieval and generation as separate systems and score groundedness over vibes.
Flaky selectors break on every UI change. A practical Playwright locator strategy layered with self-healing fallbacks that turns redesigns into non-events.
How I automated ~90% of root-cause analysis at Motive using Claude Code and MCP, cutting a 3-hour ritual down to 20 minutes of human attention.
Your agent passes every functional test and can still be hijacked by a sentence in a web page it reads. An SDET intro to adversarial agentic testing.
Your assertion library assumes deterministic output. A practical framework for testing LLM features that never return the same answer twice.
MCP made agent integrations easy and opened a supply chain. Poisoned tool descriptions, indirect injection, and the new attack surface QA must evaluate.
What separates a throwaway test-generation prompt from a reusable system that catches edge cases — and how to build the latter in an afternoon.
Using a language model to score language model outputs is clever but fragile. The failure modes, calibration strategies, and when human review still wins.
A practical architecture for Java API test suites: REST Assured for HTTP, JUnit 5 for orchestration, and Page Object Model for long-term maintainability.