Long-form notes on building and testing software with agents in the loop.
Agentic QAI ran Playwright's three test agents against this live site. The Planner found a real bug, the Generator wrote a test that encodes it, and the Healer showed why every heal needs a human review.
AI ToolingAlmost every agent failure I hit this year was a harness problem, not a model problem. What a harness is, what LangChain gained by changing only the harness, and three failures from building this site.
TestingAI agents now write, run and repair a growing share of our tests. At the base of the pyramid they make it stronger. At the top they can rebuild the ice-cream cone faster than any team ever could.
Most teams lean on slow UI tests and write few unit tests. How the testing pyramid, shift-left testing and parallel execution turn that around.
Five truths reshaping how modern teams architect quality, from Chernobyl's control room to 72-hour soak tests: safety-critical UI testing, ATDD's comeback, reusable auth state, load vs. stress, and why short tests miss the worst bugs.
A green suite isn't a safe release. Five counter-intuitive testing truths: performance engineering over load tests, conversation over ATDD tooling, reusable Playwright auth, E2E in control rooms, and spike plus soak.
Design patterns aren't a dry catalog to memorise. Five counter-intuitive lessons on managing complexity without over-engineering, from lambdas replacing Strategy to Command-powered undo.
