AI Cost for Agents
Agent cost is a workflow cost. Learn how to cap retries, prevent loops, and keep tokens under control.
The problem
Agent runs are where cost surprises happen: retries, tool chains, and self-reinforcing loops.
Why agents can burn budgets
- Tool retries multiply calls
- Long reasoning turns into more tokens
- Loops happen when there’s no convergence signal
Cost breakdown
Agent cost = tokens across all model calls + tokens generated by tool context + retries.
Real example
An agent repeats a tool call sequence because the result never satisfies the stopping condition.
Optimization guardrails
- Cap retries and depth
- Add a convergence rule (stop when you reach the goal)
- Set per-agent budgets and alerts
Quick checklist
- Stop runaway loops
- Use routing for simple steps
- Track agent cost by run