Shipping an LLM app is the easy part. Knowing why it gave a wrong answer at 2 a.m. on a Tuesday is the hard part. Traditional software fails loudly with a stack trace. LLM apps fail quietly: the code runs, the API returns 200, and the answer is confidently …

Have You Ever Asked a Friend "Why Did You Do That?" and Gotten a Shrug?

That's basically what it feels like to debug an AI agent system.

You watch it mess up. You check the logs. The logs cheerfully tell you what happened — it called a tool, it got a …