Where the Tokens Go: Heat, Entropy and Error in Agentic Coding

Impact 7 arXiv cs.AI

Summary

  • Researchers introduce a software thermodynamics theory to measure token costs in agentic coding by dividing token usage into work and heat.
  • About 80 percent of tokens in a successful run are classified as heat, while the free energy of a change is estimated at 14 to 22 percent of the mean cost.
  • Testing cycles of coding and uncoding show that software temperature rises with the sampling temperature of the model.
  • Test suites act as imperfect software Maxwell's daemons, where verdicts varying by environment incur additional token costs.

Why it matters

Applying thermodynamic principles to agentic coding introduces a quantitative framework for analyzing token efficiency and optimizing AI-driven software development workflows.

Who should care