Brief

Developers can significantly reduce token waste in AI agent tool schemas by removing unnecessary fields like "title", "$schema", and "additionalProperties". These fields, often comprising up to 20% of a schema's size, do not aid the AI in selecting the correct tool. Code snippets in Node.js/TypeScript and Python are provided to automatically strip this cruft, leading to substantial cost savings on each agent turn. AI

Two recent analyses highlight significant inefficiencies in how AI agents handle token costs, particularly concerning the data sent to language models. The first, by Zied Mnif, reveals that AI agents often resend extensive system prompts and tool schemas with every request, leading to token overhead that can be many times larger than the actual user query. The second, by Layzer Zero, introduces a GitHub project called Headroom that compresses tool outputs, logs, and RAG chunks before they reach the LLM, claiming reductions of 60-95% in token usage with minimal impact on answer quality. These findings suggest that current agent architectures may be overspending considerably on input tokens, with potential monthly savings of thousands of dollars for large-scale operations. AI