Your AI Agent Is Burning Tokens. Do You Know How Many?

Short summary

Your Claude Code sessions may waste tokens on simple tasks loading unnecessary configs. The author built a 15-line Python tracking system using edit counts as a proxy for token spend, then layered decision thresholds (L0-L3) that match your confidence to decision risk: collect at 1-4 sessions, flag outliers at 5-14, compare session types at 15-29, and project savings at 30+. The framework applies to any metric tracking needing to answer 'when do I have enough data to act?'

• •Built low-overhead token tracking system using edit count as proxy metric, validated within 30% of actual consumption across 10 manual cross-checks
• •Created L0-L3 decision framework matching data collection volume to decision confidence: collect (1-4), flag outliers (5-14), compare categories (15-29), project forward (30+)
• •Methodology is platform-agnostic — applies to cost, quality, speed, accuracy tracking, or any system trying to detect and optimize its own waste

Generated with AI, which can make mistakes.