Budgets

Budgets give you predictable ceilings on estimated spend and token volume before agents run. They are implemented in application/internal/cost/budget.go and applied in the V3 pipeline ahead of agent execution, so a run can stop or ask for confirmation instead of drifting past what you consider acceptable.

The YAML block below is the full shape: currency default, per-run and per-task caps, a daily aggregate cap, and policy flags that control blocking versus explicit override.

Schema

budgets:
  default_currency: EUR
  per_run:
    max_estimated_cost: 1.00
    max_estimated_tokens: 500000
    require_confirmation_above_cost: 0.20
  per_task:
    max_estimated_cost: 0.30
    max_estimated_tokens: 150000
  daily:
    max_estimated_cost: 10.00
  policies:
    block_when_over_budget: true
    allow_override_with_flag: true
    override_flag: "--allow-over-budget"

CLI interaction

These flags interact with the same numbers and policies: tighten a single run, exit after estimation only, or bypass a block when policy permits. --yes is the escape hatch after BudgetPendingConfirmError when you have already reviewed the estimate and want to proceed.

Limitations

Estimates multiply heuristic token counts by the rates in your pricing sheet. Provider invoices can diverge from that model, especially when vendors change tiers or bill by factors AgentFlow does not model.

Related

• Cost-aware workflow
• Token estimation