HomeGlossarySpend cap

Spend cap

Also called: hard stop, budget cap

Definition

A spend cap is a limit on what an agent can spend on model usage, and a useful one halts execution at the limit rather than sending a notification.

In practice

The distinction between a cap that stops and an alert that warns is the whole point. An alert is a report about money already spent, and it arrives while the mechanism that spent it is still running, which is fine when a person is watching and useless for an agent running over a weekend. Prepaid balances enforce this structurally: zero is a real floor, so the worst case is a paused agent rather than an invoice. Overruns almost never come from high per-call prices. They come from retry loops, context that grows every run, and tasks that fan out over a longer list than expected.

Example

An agent summarises new support tickets and normally costs a few dollars a week in model usage. One Friday a ticket with a huge attachment makes it retry the same step all weekend. With only an alert, the owner gets an email on Saturday and a large bill on Monday. With a spend cap that halts at the limit, the agent pauses when the cap is reached, and the worst case is a missed weekend of summaries.

Common questions about spend cap

What is the difference between a spend cap and a budget alert?

A budget alert reports money already spent and lets the agent keep running. A spend cap halts execution at the limit, so the worst case is a paused agent rather than a surprise invoice. For an agent running overnight or over a weekend with nobody watching, only the cap actually limits the damage.

How should I size a spend cap for an AI agent?

Size a spend cap against a bad day, not an average one. Estimate the normal monthly cost, add generous headroom, often around 50 percent, and set the cap where hitting it stops the agent. Overruns usually come from retry loops, growing context, or tasks fanning out, not from steady use.

Read more

Nearby terms

All 25 terms

Run agents without the homework

Qoren runs OpenClaw, Hermes, and Codex agents in managed environments, with the restarts, secrets, schedules, and spend caps handled for you.