How do I stop an LLM agent from running up costs?

#The checklist

1. Step cap: a maximum number of model calls per run, enforced in the loop itself.
2. Output cap: set max_tokens (or the provider's equivalent) on every call so one reply cannot run long.
3. Run budget: read the run's spend so far from your usage ledger and refuse the next call when it would pass the budget.
4. Loop detection: if the same tool is called with the same arguments twice in a row, stop or change strategy.
5. Wall-clock timeout: a run that takes ten times longer than normal is a bug, not a slow job.
6. Tenant kill switch: a daily spend ceiling per account that blocks new runs and alerts you.
7. Alert on the tail: watch the most expensive runs, not the average.

#Why cost grows faster than steps

If each step adds about the same number of tokens and the full history is sent every time, input tokens across n steps add up to a triangle, roughly n times n plus one over two times the tokens per step. Fifty steps of 1,000 added tokens each is 1,275,000 input tokens, and a hundred steps is 5,050,000, about four times as many. This is arithmetic, not a benchmark, and caching or summarising history lowers it.

The full write-up, with a small budget guard in Python, is in Runaway LLM Agent Costs: Budgets, Step Limits and Kill Switches.