Field note: agents in production make cost an engineering problem

A pattern from our client work, recorded because it keeps repeating.

Demos are free. Production is metered. Every AI agent we have shipped into a client system eventually hit the same wall: the thing works, the client likes it, and then the invoice for inference arrives and someone asks whether the workflow is worth what it burns.

So token efficiency stopped being an optimization and became part of our engineering standard. We treat model calls the way infrastructure teams treat compute: measured per workflow, budgeted per task, reviewed when the numbers drift. Shorter context where the task allows it. Cheaper models where judgment is not the bottleneck. Caching where the same question arrives twice.

Two things follow from this, and both changed how we operate.

For clients, cost discipline is part of the deliverable. An agent that solves a problem at a price the workflow cannot carry is not a solution, it is a liability with good manners. The engagements that renew are the ones where the economics were engineered, not discovered.

For us, it is a filter on what deserves to become a product. A workflow that only makes sense at demo-day prices will not survive contact with a budget. The repeated client problems worth productizing are the ones that stay worth solving after you count the tokens.

Small lesson, learned on billable time.