There is a specific moment most developers hit when building AI chains, usually somewhere around the third or fourth iteration: the demo works perfectly, the test runs fine, and then something in production produces output that is technically valid but structurally wrong in a way that breaks everything downstream. The chain was brittle. You just did not know it yet. What follows is a collection of failure patterns that come up repeatedly when developers actually ship AI automation — not the patterns covered in documentation, but the ones discovered after the documentation runs out. The JSON Problem Is Not What You Think It Is Every developer building AI chains eventually learns to prompt for JSON output. What takes longer to learn is that "prompt for JSON" and "reliably receive JSON" are different things, and the gap between them is where most chains fail. The obvious failure is malformed JSON — unescaped quotes, trailing commas, truncated output. These are catchable with a try/catch and a retry.…