Notes from the engineering trenches.

Most 'AI chatbot' tutorials stop at a prompt and a chat window. This post picks up where those leave off: how to design a support agent that survives contact with real users, real tickets, and real edge cases.

We start with the boring but critical work — mapping the conversation space, defining tools with clear contracts, and deciding what the agent is explicitly not allowed to do. Then we layer retrieval-augmented generation on top of a curated knowledge base, using hybrid search and re-ranking to keep answers grounded.

Finally, we wire in evaluation. Every deployed agent needs an eval harness: a set of scenarios that run on every change, scoring correctness, tone, and escalation behavior. Without it, you're shipping vibes. With it, you can iterate confidently and catch regressions before customers do.

The fastest way to waste an AI assistant is to paste a stack trace and ask 'what's wrong?'. The fastest way to get value is to treat the model like a sharp collaborator who just walked in: give it context, a hypothesis, and a clear question.

This post covers the patterns I use every day — reproducing the bug in the smallest possible surface, feeding the model the exact frames that matter, and asking for three plausible causes ranked by likelihood. It also covers the anti-patterns: over-trusting confident-sounding fixes and losing hours to hallucinated APIs.

Done right, AI debugging compresses hours of grinding into minutes of focused reasoning. Done wrong, it's a very expensive rubber duck.