AgentFlow
Workflows

Cost-aware workflow

Estimate, optimize context, and cap spend before agent execution.

Cost-aware workflow

This page operationalizes spec-doc §11.4: every flag and configuration key referenced here exists in AgentFlow builds. The sections walk through estimating spend, tightening context packs, enforcing budgets, and preferring local execution when routing policy allows—each step spelled out below with verbatim commands instead of abbreviated prose.

1. Per-task estimate

agentflow estimate billing-v2 --task task-003 --dry-run

Shows heuristic input/output tokens and estimated cost from your pricing sheet.

2. Context optimization

agentflow context billing-v2 --task task-003 --optimize

Builds a reduced context pack before large agent prompts. Disable reduction with agentflow work ... --no-context-reduction.

3. Run with budget

agentflow work "develop billing-v2" --budget 0.50 --estimate-only
agentflow work "develop billing-v2" --budget 0.50

Config budgets (budgets.per_run.max_estimated_cost) apply when CLI --budget is zero. Over-budget runs block unless:

  • confirmation prompt accepted, or
  • --yes after BudgetPendingConfirmError, or
  • --allow-over-budget when budgets.policies.allow_override_with_flag is true

4. Local preference

agentflow work "develop billing-v2" --prefer-local --no-cloud

--allow-cloud explicitly permits cloud routing when policies would otherwise restrict.

Explainability

work and estimate print step summaries (routing choice, context bytes, budget checks). See Token estimation.