An agent that cannot see what a call cost will spend what it has. The fix is in the API, not in the prompt.
Most AI products meter their cost somewhere a developer cannot see it at the moment it matters: a dashboard, a monthly invoice, a usage page refreshed every hour. The agent making the calls sees nothing at all. So it makes the next call.
When we built Gistly we put the price in the response. Every search returns what it charged and what the wallet holds now, both as headers and inside the body's meta. A tool that reads the body gets the cost for free; a tool that only reads headers gets it too.
Why it changes behaviour
An agent with a budget can be told to stop when the balance drops under a threshold, to prefer the quick depth until a question proves hard, and to escalate to a deeper search only when the first answer cites too little. None of that is possible when cost is a number someone looks up later.
Why it changes trust
A developer who sees the charge next to the answer stops asking whether the meter is honest. A prepaid wallet with a fixed price per depth makes a bill predictable before the first call. Pricing is part of the product's interface, not a page on the website.
The rule we apply
If a call costs money, the response says how much. If a call can be made cheaper, the request has a dial for it. If the dial changes the result, the documentation says what changes. We apply that to our own products first and to the ones we build for others next.