Running agents in production

Where the tokens actually go, why tool definitions are charged on every request, and the three levers that change the bill.

A minimal eval harness: twenty cases, a deterministic check where possible, and a rule for when to re-run it.

What to hand an agent, what to keep behind a human, and how to reconstruct what it did afterwards.

Eight checks that catch the failures which only appear once real people and crawlers arrive.