Budgets, cost per run and pause-and-ask

Where agent cost comes from

Each turn of the loop is a call to a model, and every call costs in proportion to the tokens read and written. A run costs more when it takes more turns, reads big tool results or uses a stronger model. Multiply by how often it runs and you have the monthly cost. The three levers are fewer turns, smaller context and a cheaper model where quality allows.

Know the cost before you scale

Do not guess the cost of an agent. Measure it.

• Per run. Every run header shows the steps, tokens and cost.
• Typical cost. On the agent page, the Budget panel shows what a run usually costs, with a range, based on recent runs, or estimated from model prices when there is no history yet. Estimates are not promises.
• Across all agents. Usage shows runs, tokens and cost per day and per agent.

For anything scheduled, multiply typical cost by runs per month and compare with your monthly budget.

The two budgets

• Per run (USD) is a hard stop. A run that reaches it ends with a plain failure message.
• Per month (USD, 0 = no limit) protects the month. The Budget panel shows how much has been used against it.

You are also warned as spending climbs: notifications arrive as an agent passes the halfway mark, most of the way, and the full budget.

Pause-and-ask at the monthly budget

When a run would go past the monthly budget, it does not fail silently. It pauses, and an approval appears in Approvals titled " reached its monthly budget". You choose:

• Allow this run lets this run finish, without changing the budget.
• Raise this month by adds the amount you enter to this month's budget only.
• Decline ends the run as Cancelled.

The run waits in the status Waiting for your budget decision (on a task) while it is on hold. If nobody decides, it eventually expires. Test runs and replays still fail at the budget rather than asking.

Steps to try

1. Open an agent and look at the Budget panel. Read the typical cost line and the monthly bar.
2. Lower Per month (USD, 0 = no limit) to a small amount so that your next run will pass it. Save.
3. Run the agent. Watch it pause, then open Approvals and read the budget card.
4. Choose Allow this run. The run continues. Then restore your monthly budget.
5. Open Usage and find the run's cost.

Good habits

• Set a per-run budget close to a few times the typical cost, not hundreds of times.
• Review Usage weekly while an agent is new.
• Remember that your AI provider's invoice is the true amount. Figures in PGB Agents are estimates from token counts.
• If an agent keeps hitting its budget, find out why in the traces before raising it.

Recap

• Cost follows turns, context size and model choice, multiplied by frequency.
• Per-run budgets are a hard stop; the monthly budget pauses the run and asks you.
• Choose Allow this run, Raise this month by or Decline.
• Measure with the run header, the Budget panel and Usage before scaling.