AI
Where the Tokens Go: Heat, Entropy and Error in Agentic Coding
Publication: cs.AI updates on arXiv.org
Author: Al\'an Aspuru-Guzik, Varinia Bernales
Source published: 19/11/2026, 05:00:00 UTC
Briefing created: 19/11/2026, 05:13:05 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 19/11/2026, 05:03:24 UTC.
Briefings
A new theory proposes that coding agents consume tokens as energy, where successful edits generate 'heat' (tokens staying in place) while failures dissipate energy. The study tests this by running agents that undo their changes, measuring token costs against model sampling temperature. Approximately 80% of tokens in successful runs are classified as heat, and successful coding costs 1.1 to 1.5 times more than uncoding.
Key points
- Agents consume tokens as energy, with successful edits generating 'heat' (tokens remaining in place) and failures dissipating energy.
- Approximately 80% of tokens in successful runs are classified as heat, while successful coding costs 1.1 to 1.5 times more than uncoding.
Context
This research applies classical and stochastic thermodynamics to agentic coding, treating token usage as energy cost. The study uses a Maxwell's daemon test suite to measure energy dissipation during agent cycles, offering a thermodynamic framework for analyzing coding efficiency and agent behavior.
Limitations
The study relies on a single test suite whose verdicts vary with environmental conditions. The free energy estimate uses the Jarzynski equality approximation, which may not capture all energy dissipation mechanisms. Uncertainty exists regarding the precise relationship between sampling temperature and heat generation across different agent configurations.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 19/11/2026, 05:13:05 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeicslnuf6p7vpfmvfl5tjblm5gbgyibsoh3cv7vs573lchivvukeke