How the run budget is counted
The budget applies to the run as a whole rather than to each individual model reply. Tool calls contribute when they report a credit cost, and a subagent uses the same shared budget as its parent. This makes the limit useful for controlling a multi-step task whose cost includes research, file work, and other tools.
The agent checks the amount spent between steps. A single step that is already underway can bring the total to or beyond the configured threshold; the limit prevents the next step from starting. For that reason, treat the budget as a stopping threshold, not a promise that the final total can never cross it by the cost of one operation.
Choose a limit that fits the task
Give a small budget to a narrow lookup or a first-pass review, and allow more room when the agent must inspect many sources or iterate on a complex change. Pair the budget with a sensible step limit: the budget controls credit use, while the step limit controls how many rounds the agent may attempt.
If a run stops for budget, review the work and outputs already produced before raising the limit. You can continue with a focused follow-up that names the remaining work, or start a new run with a larger budget when the unfinished steps justify the added cost. A budget does not make a task succeed automatically; it gives the run a defined spending boundary.