A demo is a controlled environment. Production isn't.
When you demo an agent live, you are quietly doing a lot of invisible work. You pick inputs you know it handles well, you rephrase a question if it stumbles, you interpret a slightly-off answer generously, and if something goes sideways you just try again off camera. None of that safety net exists once the agent is running unsupervised against real users typing whatever they want, at 2am, with typos, ambiguous requests, and edge cases you never thought to test. The agent's core reasoning ability hasn't changed between the demo and production. What changed is that nobody is there to catch it when it drifts.
This is why teams are consistently surprised when a demo that looked 90% reliable turns into something that fails constantly in the field. The 10% you didn't see in the demo isn't a rounding error, it's the long tail: malformed inputs, tools timing out, users asking for something the agent was never scoped to do, two steps of a plan silently contradicting each other. Production readiness is almost entirely about how gracefully the system handles that long tail, not about making the agent smarter.