How do I keep an AI agent from running up a big API bill?
Set limits at the provider
Most model providers let you set monthly spending caps and usage alerts in the billing dashboard. Set a soft alert well below your cap so you get warned early. Check the exact steps on each provider's billing page, since menu names change over time.
Create separate API keys for each project as well. If one agent misbehaves, you can revoke only that key without stopping everything else you run.
Limit steps and tokens inside the agent
Many agents overspend because they keep calling tools without a clear stop condition. Set a maximum number of steps per task, a maximum output length for each model call, and a timeout. When a limit is hit, the agent should stop and report what it finished.
Shorter context also reduces cost. Trim old messages and summarize long tool outputs before sending them back to the model.
- Cap steps per task and per hour
- Set a maximum token limit on each model call
- Trim tool output before adding it to the context
- Send simple steps to a cheaper model when quality allows
Track what each run costs
Log the model name, input tokens, and output tokens for every call. A spreadsheet or a small logging library is enough to show which task costs the most. Once you know where the money goes, you can fix the expensive step instead of guessing.
Common mistakes
- Setting one high spending cap and assuming it protects you from every problem.
- Letting an agent retry failed tool calls with no limit on the number of retries.
