Moving a sales agent to a new production runtime involved more than choosing the runtime. We verified the connection to the team’s work surface, preserved pending workflow state, checked approval ownership and duplicate controls, tested restart behavior, and restored a backup into an isolated location to prove recovery.
The old runtime stayed available as a rollback while the new release was observed. No test message reached a customer. These are documented delivery tasks, not estimates of the client’s costs.
They show why model usage is a poor proxy for the price of a dependable system. Integration, state, testing, recovery, and continued care are real work even when the visible result is a short conversation.
From our article What an AI agent really costs.