Budgets and oversight

Three things make agent work answerable: you can see what it costs, cap it, and stop it.

Usage

Settings → Usage shows, for any period: tokens and cost per day, and a breakdown by agent, by model and by person. Costs come from the prices on each model (Models).

Delegated work is attributed to the agent that did it, so a team of agents shows where the money actually went.

Budgets

Settings → Budgets sets a monthly cap, per UTC month, on one of three things:

• everything this owner spends;
• one agent — its own calls, including when another agent delegates to it;
• one team — what its members' conversations and tasks cost, on any agent.

Each budget has an alert level, 80% by default. Crossing it notifies the owner's managers and the agent's creator, once per level per month. When the budget is used up, new runs are refused with a message naming the budget — answers in progress finish.

Budgets are checked before a conversation run, before a task run, and before each delegation.

A budget is a cap on spending, not a security boundary. It stops a runaway agent from costing you a fortune; it does not stop anyone from doing anything they are allowed to do.

The audit log

Settings → Audit log records what happened in this owner: sign-ins and account changes at the platform level, and here, every agent published, shared, archived or deleted, every setting changed, every tool and integration call with its input, and every transcript a manager read.

Filter by action, by person, by date, or search. Each entry has who did it, when and to what.

Owners see their own log; the platform's administrators see sign-ins, accounts and organizations being created — not what is inside any of them.

Stopping an answer

Stop in a conversation ends the answer being written, including anything it has delegated. The partial answer is kept, and the run is recorded as stopped with who stopped it. The managers of an agent's owner can stop a run on it as well as the person who started it.