Cost control is the difference between running an AI-native company and running an expensive science experiment. Company Agents enforces budgets at six nested scopes, captures every token, and gives you a live cost panel that shows where the money is going before it is gone.

​
The six budget scopes

Budgets stack. A single agent action has to fit inside six envelopes at once:
  1. Company budget — the top line. Every cost in the company rolls up here.
  2. Team budget — the share allocated to a team for the current period.
  3. Agent budget — the share allocated to one agent within the team.
  4. Workflow budget — a ceiling on the workflow definition itself (so a badly written workflow cannot run wild).
  5. Task budget — the ceiling on one specific task run.
  6. Loop budget — a hard cap on any single inner loop inside a task, to catch runaway iteration.
Every action has to be affordable under all six. When the tightest one blocks, the agent stops and the orchestrator records a budget breach.

​
What gets counted as cost

The orchestrator captures three kinds of cost:
  • LLM tokens, priced per model per 1K tokens, at the rates configured under Settings → Pricing
  • Tool-call costs (external APIs the agent used, metered per the integration’s own pricing)
  • Runtime cost (infra time: minutes of web preview, minutes of a crawler session, gigabytes of storage)
All three are aggregated by task, run, agent, team, and company, and are visible in the cost panel. Non-LLM costs are often invisible in other platforms because nobody meters them. Company Agents defaults to on.

​
Reading the cost panel

The cost panel lives at Dashboard → Cost and has three views:
  • Live — spend in the current hour, minute, and current run. This is what you watch when something is running.
  • Period — spend across the current fiscal period (day, week, month) with breakdowns by team, agent, and workflow.
  • Forecast — projected spend to end of period based on recent run rate, with a confidence band.
Every row drills down to the individual tasks and runs that contributed. You can always trace a dollar back to the action that spent it.

​
Setting a budget

When you create a company, you set a company-level budget and a fiscal period. The system suggests a default team-level split (equal shares across teams, 80% of the company budget, 20% held in reserve). You can override the split at any time under Company → Budgets. Team-level budgets cascade down to agents the same way. Workflow and task-level budgets are set on the workflow or task itself. Loop budgets are set once under Settings → Safety → Loop budget and are global per adapter. The default is aggressive (about one hundred tool calls per loop) so that stuck agents fail fast.

​
When a budget breaches

Breaches come in three flavors:
  • Soft breach — the agent is running low. The orchestrator sends a warning event and the agent is asked to tighten its own plan.
  • Hard breach — the agent is blocked from starting its next action. The task pauses and a human is notified.
  • Catastrophic breach — spend rate over a very short window (default: 5x the expected rate for the task class). The orchestrator kills the process group immediately and flags the run.
All three are visible in the cost panel and in the activity log. You can configure the thresholds for each under Settings → Safety → Breach thresholds.

​
Top-ups and reserves

If a team is about to breach and the work is important, you can top up the team’s budget from the company reserve (the 20% held back at company creation). Top-ups are logged and shown on the cost panel so there is no ambiguity about why a team went over its original allocation. You can also set auto top-ups: rules that automatically move reserve into a team when the team crosses 80% of its period budget and is on a workflow tagged as critical. Auto top-ups have their own hard ceiling so they cannot drain the reserve without a human in the loop.

​
Why stacked budgets matter

A single global budget is easy to breach in subtle ways. One agent can eat the whole month’s spend in an afternoon if nothing stops it at the team or task level. Stacked budgets force the math to work at every scope. The six envelopes together mean:
  • A single runaway task cannot drain the team
  • A single runaway team cannot drain the company
  • A single runaway loop cannot even drain the task
  • Every dollar has a clearly attributable owner
This is the shape of cost control that survived contact with real production use.

​
Next

  • Approvals for how spend thresholds trigger human-in-the-loop gates.
  • Activity log for where breach events are recorded.
  • Cost reporting if you are the one writing an agent and want to understand how to attribute costs from inside an agent run.