Three weeks after launch, finance forwards the invoice with a one-line email: is this right?
Nobody on the team can answer.
Not because the number is secret. Because it is one number, and the thing that produced it was not one thing.
We had latency budgets. Eval suites. Retry logic. Dashboards for all of it.
The one decision that drove both quality and cost the most, which model runs which step, was vibes.
We build agentic systems for a living. Here is what that cost us.
