80% of coding agent tokens are wasted as heat A paper posted to arXiv (2611.23375) reports that 80% of tokens in agentic coding runs are wasted as "heat" — model calls that do not change the code but still consume budget — making per-task token cost 5–7x higher than the actual work done. The paper attributes the inefficiency in part to sampling temperature and estimates that optimizing for the 14–22% of tokens that advance a task could cut cloud spend by 4–6x or let agents handle 5x more complex repositories on the same budget. arXiv https://arxiv.org/abs/2611.23375 80% of coding agent tokens are wasted as heat Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. 80% of tokens in agentic coding runs are wasted as "heat"—model calls that don’t change the code but still burn budget. This means your per-task token cost is 5–7x higher than the actual work done, and sampling temperature directly drives that inefficiency. Optimizing for thermodynamic "free energy" the 14–22% of tokens that actually advance the task could cut your cloud spend by 4–6x or let you ship agents that handle 5x more complex repos on the same budget. 80% of tokens in agentic coding runs are wasted as "heat" on model calls that don't change the final output, meaning most LLM compute is spent on unproductive exploration. This reveals a direct optimization lever: reducing sampling temperature or constraining exploration could cut token costs by 4–5x without impacting correctness, making agent workflows far cheaper to run in production.