Most "AI agent" demos work because exactly one person is using them — usually the person who built them.
Production is different. Real users send malformed inputs, the API rate-limits, the model picks the wrong tool, the vector store returns stale results on day 90, and somebody asks for a feature your prompt scaffold can't bend around.
Half my client work right now is turning agent prototypes into things that survive contact with actual users. The unsexy parts — retries, idempotency, eval suites, observability, structured tool I/O — are 80% of the real build.
If your agent works in the demo and breaks in prod, the demo wasn't the product. The retries were.
