Guide · Agent orchestration
Set a Budget for Multi-Agent Work Before You Start It
Budget for the whole task, including worker calls, retries, tools, and the time needed to review the result.
A run with several agents has several places to spend money. The coordinator consumes tokens, workers consume tokens, tools may charge separately, and retries can repeat work you thought was finished. Set a budget for the completed task before deciding how many workers to allow.
Estimate one representative job
Start with a small trial and record the actual usage of each step. Include the coordinator's input and output, worker calls, paid tools, and repeated attempts. Keep elapsed time and human review time next to the monetary cost so the comparison reflects how the work will be used.
For a hypothetical supplier review, the unit is one accepted review of a defined set of documents. It is not one model response. If the first response needs another research pass and substantial correction, those costs belong to the same job.
Record the model and settings used by each worker. GPT-6 Astra and Claude Fable 5.1 may occupy different roles in your workflow, and changing a role's model can change both its usage and its review burden. Check current provider pricing when calculating costs instead of embedding a remembered rate in the budget.
Put limits at useful boundaries
A total spending cap is helpful, but it should not be the only control. Limit concurrent workers, repeated attempts on the same failure, and the amount of work an agent may create for other agents. Decide what should happen when a task approaches its limit.
Claude Managed Agents documents a shared session budget across threads. Other platforms may expose different controls. Confirm what your chosen runtime counts, and keep application-side limits for costs or external services that fall outside the provider's budget mechanism.
- Continue only the work needed to preserve the current result.
- Stop creating new assignments when the remaining budget is too small.
- Return accepted artifacts and unresolved items when the run pauses.
- Require an explicit budget decision before restarting expensive work.
Make a budget stop useful
A stopped run should leave a readable account of what finished, what remains, and what is needed to continue. Save completed research and patches as they are accepted. Otherwise, a budget limit can force the next run to pay for the same discovery again.
Avoid making the final summary the only place where useful work becomes durable. Write accepted artifacts during the run, and keep their identifiers in the task record. That makes a partial result useful even when the coordinator never reaches its planned final response.
Revisit the budget from completed work
Compare similar tasks over time and look at the expensive outliers. A repeated retry loop or duplicated research may explain more than the chosen model's rate. Fix that behavior before assuming the answer is always a cheaper model or a larger cap.