concepts

Limits, budgets, and guardrails

Tell mechanical execution ceilings, governed spend, and policy stops apart when a responsive agent cannot continue.

RuntimeintermediateCurrent releases
Verified 2026-08-25View sourceReport a docs issue

The model is responding normally when the Run stops. “The AI crashed” is an attractive explanation—and often the wrong one.

The tenth loop step

Invocation and loop limits cap how many model or execution cycles one attempt may consume. Timeouts and output bounds protect a specific boundary. Reaching one should produce evidence naming that ceiling rather than a generic provider failure.

Limits keep a malformed or unproductive loop finite.

The third identical attempt

No-progress detection addresses a different problem: the Run is spending steps without changing its situation. It can stop below a numeric ceiling because repetition itself is the evidence. The remedy is a different plan, not simply a larger loop limit.

The last authorized dollar

Budgets represent owned resources such as model cost or another metered allocation. A durable ledger can reserve value before dispatch, commit actual spend, release unused reservation, and reconcile uncertain outcomes. Idempotent allocation evidence prevents a retry from charging the same logical use twice.

Budget exhaustion is not a model capability failure. Increasing a token limit does not replenish an organization allocation.

The action policy forbids

Guardrails stop or redirect work because it violates an execution or autonomy rule: disallowed tool class, missing capability, approval requirement, project restriction, or another policy-owned boundary. They should identify the owning rule and the action that could not proceed.

Depending on the stop, the Run may fail, wait for an exact intervention, or preserve valid accumulated history for a later attempt. It must not report success because partial work looked promising.

The remedy starts with the owner

Operators need these categories to choose a real remedy: reduce scope, increase an approved limit, allocate budget, change the plan, satisfy an approval, or fix configuration. A blanket retry does not answer which premise changed.

Failures, retries, and recovery deals with what can safely happen next. Operational budget ownership is documented in Budgets and allocations.