The Hidden Costs of AI Agents: Retries, Loops, and Tool Calls
Agents multiply cost through retries, tool chains, and non-converging loops. Learn to measure and cap the runaway parts.
An agent that calls a model five times per user message bills five times the tokens—before you add retries when tools fail.
Where agent cost hides
A single user action in an agent can trigger:
- Multiple model calls (plan → act → observe → refine)
- Tool calls that return verbose results
- Retries when a tool fails or times out
- Loops that never reach a stopping condition
Each step adds billed tokens on top of the previous ones.
The retry trap
A tool fails 5% of the time. You retry up to 3 times. Suddenly call volume is higher than it looks, and so is cost—even when average tokens per call stay flat.
Non-converging loops
If an agent has no clear "done" signal, it keeps refining. The loop's cost is unbounded until something stops it.
How to cap the runaway parts
- Retry caps: limit attempts per tool
- Depth caps: limit how many reasoning turns are allowed
- Convergence rule: stop when the goal is met, not when it "feels" done
- Per-agent budgets: hard stop at a spend threshold
Quick checklist
- Count calls per user action, not just tokens
- Cap retries and loop depth
- Set per-agent budgets and alerts
Related reading
Set up AI API cost monitoring that catches token waste, retry spikes, and runaway agents before they hit your bill.
Batch APIs offer large discounts for asynchronous jobs. Learn when to route work to batch and when real-time is worth the premium.
Model routing sends simple tasks to cheap models and reserves premium models for hard ones—often cutting cost without losing quality.
Estimate AI API cost from tokens and request volume before shipping, so you don't discover unusable unit economics after launch.