Building CodeReviewr, the team found their LLM agent reading 47 files when a simple grep would suffice. Prompt engineering and fixed limits failed to curb the token-hungry behavior. The solution was a credit-based budget system, assigning costs like 1 credit for grep and 10 for reading files.

Starting with a 100-credit budget per task forced the model to prioritize cheaper tools. An extend_budget tool required specific justifications, with 90% of requests deemed legitimate. This led to a 45% reduction in tool calls across 100+ code reviews without any drop in accuracy.

The team is now exploring dynamic budgets based on task complexity. This approach mirrors how humans manage scarce resources, proving more effective than vague instructions. For developers building tool-using agents, implementing a budget is a practical way to control costs and teach strategic thinking.