{"slug": "80-of-coding-agent-tokens-are-wasted-as-heat", "title": "80% of coding agent tokens are wasted as heat", "summary": "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.", "body_md": "[arXiv](https://arxiv.org/abs/2611.23375)\n\n### 80% of coding agent tokens are wasted as heat\n\nWhich summary reads better? Pick one — models revealed after.Both summaries are AI-generated.\n\n80% 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.\n\n80% 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.", "url": "https://wpnews.pro/news/80-of-coding-agent-tokens-are-wasted-as-heat", "canonical_source": "https://www.snipvote.com/story/cmwn879q00002b5bg86ukjrey", "published_at": "2026-11-19 07:43:14.081272+00:00", "updated_at": "2026-11-19 07:43:16.879637+00:00", "lang": "en", "topics": ["ai-agents", "large-language-models", "ai-research"], "entities": ["arXiv"], "also_reported_by": [], "takeaway": "Roughly 80% of tokens in agentic coding runs do not change the code, and lowering sampling temperature or constraining exploration could cut token costs 4–5x without affecting correctness.", "recommendation": {"advice": "Cut sampling temperature and constrain exploration in coding agents to eliminate unproductive token spend; the paper measured 80% of tokens as wasted and a 4–5x cost reduction without correctness loss.", "playbook": "llm-cost", "playbook_url": "https://wpnews.pro/playbooks/llm-cost"}, "alternates": {"html": "https://wpnews.pro/news/80-of-coding-agent-tokens-are-wasted-as-heat", "markdown": "https://wpnews.pro/news/80-of-coding-agent-tokens-are-wasted-as-heat.md", "text": "https://wpnews.pro/news/80-of-coding-agent-tokens-are-wasted-as-heat.txt", "jsonld": "https://wpnews.pro/news/80-of-coding-agent-tokens-are-wasted-as-heat.jsonld"}}