Set budgets

Settings → AI Configuration → Budgets caps every run in your workspace. It needs the manage_model_policy permission.

A budget is not only money. It is four ceilings, because a run can go wrong in four ways: it can cost too much, take too long, read or write too much, or loop.

The tenant defaults

Field	Caps	Alpenglow
Soft cap	Spend that triggers a warning	$4
Hard cap	Spend that aborts the run	$12
Wall-clock timeout	Elapsed time per run	1800 s
Max input tokens	Tokens read per run	400,000
Max output tokens	Tokens written per run	120,000
Max iterations	Agentic loops per run	24
Throttle threshold	Fraction of the hard cap where throttling starts	0.8

Leave a field blank to inherit the system default. These apply to every run without a task-type override.

The gap between the soft and hard caps is the part worth thinking about. Too narrow and a warning is indistinguishable from an abort. Alpenglow’s 4 and 12 give a run room to finish after it has been noticed.

Run execution

Run execution sits with the budgets rather than with the AI runtime, and it answers a different question: whose runner executes new runs, as opposed to who produces completions. Inheriting the system default is right for almost every workspace.

Per-task-type overrides

The table under the defaults holds one row per task type. Click Edit to override any of the seven ceilings for that type; a dash means the row inherits.

Alpenglow overrides one type. Content runs get $1 soft, $3 hard, 600 seconds, and 8 iterations, because writing a launch post is a shorter and cheaper job than working through a code change, and a content run that has looped eight times is stuck rather than thorough.

Override a type when its work has a genuinely different shape. Overriding all seven for all seven types is a sign the tenant default is wrong.

What enforcement looks like mid-run

To an agent, a budget is not advice.

• At the soft cap, the run is warned and throttled: it keeps working with less headroom, and the warning is on the run.
• At the hard cap, the run stops. The attempt ends with the cap as its reason, and whatever it had produced is kept rather than discarded.
• At the wall clock, the token ceilings, or the iteration limit, the same thing happens, with that limit named.

A stopped run is not a lost one. It reports what it did, what it spent, and which ceiling it hit, and you can raise the budget for that run and let it continue.

A hard cap of zero stops everything

Blank inherits; zero is a real ceiling of zero. If runs start aborting immediately after a budget change, check for a zero somebody meant to leave empty.

Next steps

• Budgets: the concept, including agent-level budgets.
• The model catalog: the per-token prices these caps are spent at.
• The task workspace: where a stopped run reports the ceiling it hit.