What is Step Budget?
Also called iteration limit, action budget.
A step budget is the maximum number of actions, usually model calls or tool invocations, that an agent may take within a single run. It bounds the cost and duration of any one execution, converting a potentially unbounded loop into a run whose worst case is known before it starts.
Budgets exist because the worst case matters more than the average. A task that normally completes in eight steps might, on a malformed input or an unavailable dependency, consume hundreds. Without a cap that run costs whatever the loop happens to cost, and the only way to notice is a bill or an alert. With a cap, the ceiling is a design parameter.
Where the budget is counted determines what it actually protects. Counting model calls tracks spend closely, counting tool calls tracks side effects and external load, and counting completed subtasks tracks progress. Systems that care about both money and blast radius usually count more than one, since a single number cannot represent both.
Budgets can be exposed to the agent or hidden from it. Telling an agent how many steps remain lets it prioritize and summarize before running out, which produces more useful partial results. It can also encourage rushing. Hiding the budget produces cleaner reasoning but abrupt truncation, so many designs disclose only when the remaining budget is low.
The value should be derived from observed traces, not guessed. Recording the step count of successful runs gives a distribution, and a cap set somewhat above the tail catches genuine runaways while leaving legitimate long runs intact. A cap that trims a meaningful share of successful runs is too tight and shows up as an unexplained incompleteness rate.
Key points
- Caps actions per run, bounding worst case cost and time
- Count model calls, tool calls, or subtasks depending on the risk
- Disclosing remaining budget yields better partial results
- Set the cap from observed traces, not intuition
- Hitting the budget is an incomplete run, not a success
In practice
A support triage agent averages nine tool calls per ticket, with the slowest successful runs near twenty five. The budget is set to forty. A ticket containing a broken attachment link sends the agent into repeated fetch attempts, and the run stops at forty calls with a partial triage and a flag for review, rather than looping until someone notices the spend.