Cut agent token spend 60% with context routing

Summary

Agent loops burn tokens re-reading unchanged history and running expensive models on cheap tasks. These patterns shift cost from per-token overhead to architecture, freeing budget for longer reasoning where it matters.

Why it matters

Agent loops burn tokens re-reading unchanged history and running expensive models on cheap tasks. These patterns shift cost from per-token overhead to architecture, freeing budget for longer reasoning where it matters.

Sources

1. 1.cut costs 60% with zero quality loss
2. 2.agents were burning 40-50% of tokens re-parsing conversation history
3. 3.Moving tool definitions from the middle of prompts to the top saved ~15% per session
4. 4.70% of agent tasks are utility and planning, not creative reasoning
5. 5.Monthly API spend $410 before, $165 after
6. 6.Avg tokens per session 12,400 before, 5,100 after
7. 7.Self-healing success rate unchanged at 91%