Governance
A budget is a control too
A runaway agent rarely does anything dramatic. It just keeps going, and you find out from the invoice. Why spend belongs beside approvals in an agent's configuration.
The most common expensive agent failure is not a wrong answer. It is a loop. A tool returns something the model does not expect, the model tries again with a small variation, and it keeps trying - politely, tirelessly, at a few cents a turn - until someone notices. Usually the someone is in finance, and usually it is the following month.
We think about spend the way we think about approvals: as something the platform enforces, not something the agent is asked nicely to respect.
An invoice is not a control
Most teams' first cost control is the model provider's billing page. It tells you what happened across every agent you run, as one number, after it happened. That is accounting. A control has to do two things an invoice cannot: act before the money is spent, and act on the thing that is spending it.
"Keep your tool calls to a minimum" in a system prompt is the same category of mistake as "ask before sending anything". It is a request made to the one component in the system that cannot see its own bill.
Per agent, or it means nothing
An organisation-wide limit is a blunt instrument. When it trips, everything stops - including the agent that was doing useful work for forty cents a day - because one experimental agent on another team spent the month's allowance over a weekend.
So cost in Yekar.AI is attributed to the agent that incurred it, and the cap sits on the agent too. Each agent carries its own spend limit, with an alert before it is reached. The agent that loops runs into its own ceiling. The one next to it never notices.
A limit that stops everything is a limit nobody dares to set low enough to matter.
Prepaid means the stop is real
Model usage on the platform draws on prepaid credit rather than piling up into a bill. That changes what a limit is. With post-paid usage, a cap is a promise that someone will be warned. With a prepaid balance there is a number that cannot go below zero, and every movement against it is a line in a ledger that only ever gets added to.
The ledger matters more than it sounds. When someone asks why the support agent cost three times as much in March, the answer should be a list of entries, not a reconstruction.
Three things to decide before you switch it on
- What is this agent's work worth in a month? Set the cap from that, not from what the model happens to cost today.
- Who hears about it when the alert fires - the person who built the agent, or the person who owns the budget? They are often not the same person.
- What should happen at the limit? For most agents, stopping is right. For the few where it is not, that is a reason to raise the cap on purpose, not to remove it.
None of this makes an agent cheaper. It makes the cost a decision somebody took, with a name beside it, instead of something that happened to you.