arXiv
Language Models
Why AI Coding Agents Burn Through Tokens Like Overheating Laptops
If text generation is like a quick email draft, agentic coding is like debugging someone else's codebase while thinking out loud — it rehashes context thousands of times, burning through the budget fast.
This means if you're deploying AI agents to write or fix code, you need to budget for 1,000x more token spending than a simple chat — and now researchers have mapped exactly where that cost goes.
Bug reported: No