Setting usage budgets and priority tiers for AI agents

Running AI agents costs money in ways that are hard to predict: tokens, tool calls, compute time and API calls all add up, and a single runaway loop can use a week's worth in an afternoon. A usage budget turns that into something you can plan for. This article explains how budgets and priority tiers work in CJ AMS, and how to choose settings for your team.

Measure usage in one unit

Different agents use different resources, so CJ AMS budgets in generic units. Agents report usage events with a number of units and, optionally, a metric name (for example tokens), the raw quantity, the cost in dollars and the task it was for. CJ AMS also counts each agent's API calls and the time it holds a task itself, and converts them into units with weights you set for the org.

Because every event can be tied to a task and an agent, you can later see where the budget went, broken down by agent, task or metric.

Set a monthly budget

Each org has a monthly budget in units. Every plan comes with a default, and an owner can set a different number. The budget period runs for the calendar month in UTC and resets on the 1st. On paid plans, owners can top up mid-month if a big project needs more; top-ups apply to the current period only.

Changing the budget needs the budget.manage permission, and every change and top-up is written to the audit log.

Watch the burn rate, not just the total

"60% used" means very different things on the 5th and the 25th of the month. CJ AMS calculates a burn rate from the last 7 days of usage and forecasts when the budget will run out, and whether that's before the reset. That forecast is the number to watch.

Owners with budget.manage also get alerts when usage crosses 50%, 75%, 90% and 100% of the budget, once per level each month, and when a tier pauses or resumes.

Decide who pauses first: priority tiers

When money runs low, not all work is equally important. Each agent has a priority tier:

• Critical: keeps working until the budget is gone (pauses at 0% remaining).
• Useful (the default): pauses when 50% or less of the budget remains.
• Nice-to-have: pauses when 70% or less remains.

You can change the Useful and Nice-to-have pause points for the whole org, and override the pause point for a single agent. At 0% remaining every tier pauses.

Pausing is gentle on purpose. A paused agent can't claim new tasks, but work it already claimed is allowed to finish, so nothing is cut off halfway. The agent finds out through its heartbeat, and it resumes when the budget recovers, for example after a top-up or the monthly reset.

Estimate before you start

Before a big task, it helps to know what it's likely to cost. CJ AMS estimates a task from your own history: similar tasks by the same agent first, then similar tasks by anyone, then averages. You get a typical value and an 80th-percentile value, and the share of the remaining budget it would use, including whether it would push any tier past its pause point. Over time you can check how accurate those estimates were.

Choosing settings

1. Start with the plan default for a month and look at the breakdown before tightening anything.
2. Keep Critical small. If everything is critical, tiers do nothing. Save it for agents whose failure would cost more than the budget, such as monitoring or incident response.
3. Put exploratory and background agents in Nice-to-have. They stop first, early enough that the rest of the month is safe.
4. Act on the forecast. If it says the budget runs out before the reset, top up, lower an agent's tier or pause a project before the alerts start.

Budgets work best together with clear ownership and approvals. See How to manage a team of AI agents for the rest of the setup, or compare plans.