Notes on testing systems that don't behave the same way twice — flaky browsers, parallel workers, LLM judges, and the AI agents built to test them.https://qacraft.vercel.app/https://qacraft.vercel.app/posts/automating-an-ai-interviewer/https://qacraft.vercel.app/posts/automating-an-ai-interviewer/Three turns of chat completion in a for-loop, and a test that asserted six messages existed. Everything after that was the product refusing to stay that simple.Sun, 13 Sep 2026 12:30:00 GMT