Why AI Agents Cost More Than LLMs (And How to Stop Bleeding Tokens)

Short summary

AI agents cost roughly 1.8x more than simple LLM calls because they require multiple model turns to decide and execute tools. A real bookmark app demonstrates scaling from 1,700 tokens for direct summarization to 7,900 tokens for a 4-turn agent flow. Caching tool schemas, using cheaper models for intermediate steps, and batching tool calls reduce costs by 50-70%.

• •Agents incur an agent tax: 1.8x cost multiplier because system prompts and tool schemas are sent with every turn
• •Real benchmark: bookmark summarization scaled from 1,700 tokens (direct) to 7,900 tokens (4-turn agent), a 5x increase
• •Top optimizations: prompt caching (biggest lever), tiered models (cheaper for routing, expensive for final answer), parallel tool execution

Generated with AI, which can make mistakes.