Five ways your AI coding agent wastes tokens (and how to fix each one)

Short summary

AI coding agents waste tokens through five mechanical leaks: low cache hit rates, long session bloat, MCP tool schema overhead, wrong model selection, and poor monitoring. Fixes include maintaining stable prompts for caching, compacting sessions, disconnecting unused servers, routing to cheaper models, and tracking burn rate by session and hour. Cache reads are 10x cheaper than fresh input on Claude.

• •Five token leaks: prompt cache misses, long context bloat, tool schema overhead, oversized models, inadequate monitoring
• •Specific fixes: maintain stable prompts, compact sessions, disconnect unused MCP servers, route cheap work to lighter models
• •Cache reads cost 10x less than fresh input; track cache hit rate, context size trends, and token burn rate in real-time

Generated with AI, which can make mistakes.