Budgets

Agents cost money to think. A budget is the ceiling on that cost, resolved before a run starts and enforced while it works.

The layers

A run does not read one number. It resolves its budget through a chain, where each layer can narrow what the layer above allowed.

flowchart TB
D["Platform default<br/>sized by kind of work"]
T["Workspace budget<br/>soft and hard caps"]
TT["Per task type<br/>overrides by type"]
A["Agent budget<br/>daily and per task"]
R["Run<br/>resolved caps and limits"]
D --> T --> TT --> A --> R

Budget resolution: each layer narrows the one above it, and the run executes inside the result.

• Platform default. Every run starts from a default sized by the kind of work. Heavy work, such as building a feature, gets a larger envelope and more iterations than a small documentation change.
• Workspace budget. Your soft and hard spend caps for the whole workspace, plus optional per-step and per-task ceilings.
• Per task type. Overrides for a specific kind of task, so one type can be capped tighter than the rest.
• Agent budget. A daily limit and a per-task limit on a single agent, which is how one teammate is kept small without touching the others.
• The run itself. What the run finally executes inside.

Delegation can only shrink. When one agent hands work to another, the delegated budget is capped by the parent’s, never expanded by it.

What a budget actually contains

A resolved budget is more than a dollar figure. It carries every limit a run is judged against:

Limit	What it stops
Soft cap	The point where the run pauses and asks
Hard cap	The absolute ceiling, which a run may not cross
Wall clock	How long a run may take
Token limits	How much a run may read and write
Maximum iterations	How many implement and review cycles it may go through
Review threshold	The score a review must reach for the work to pass

Two of those are the ones people notice. The soft cap is a checkpoint: a run that reaches it stops and waits, and the pause is visible on the run. The hard cap is a wall: crossing it is not something an approval at the moment of the crossing can arrange.

Raising a limit is a decision, not a setting

When a run pauses at its soft cap, the fix is to raise the budget for that task. Raising is an authorized action: the caller has to be the project owner or a listed approver, and the raise is recorded in the audit trail with who granted it and how much they added.

An agent cannot raise its own budget. That is the whole point of a soft cap: it converts “this is costing more than expected” from something you find out afterwards into a decision somebody makes on purpose.

Budgets and trial credits are different things

During the beta a new workspace has trial credits, a granted balance of tokens funded by the platform. Budgets and trial credits limit different things, and both apply:

• Trial credits decide whether the platform will fund the thinking at all. Exhausting them stops agent work until you pair a runner or connect your own credential.
• Budgets decide how much a single piece of work may consume, whoever is funding it.

Next steps

• Runs and attempts: what a budget is enforced against.
• The autonomy ladder: the permission limit that sits beside the spend limit.
• Trial credits: the beta balance and what happens when it runs out.