Software agents switching between simple rules and complex reasoning follow the same thermodynamic laws that turn gas into liquid.
April 29, 2026
Original Paper
κ-Desktop: Information Phase Transitions in an Adaptive Three-Box Desktop Agent
SSRN · 6444378
The Takeaway
Information processing in an adaptive desktop agent is not a random sequence of events. The transition between rule-based logic and LLM-driven thought follows a predictable mathematical growth law. These phase transitions are identical to the physical shifts seen in matter as it changes state from gas to crystal. This means the behavior of complex software can be predicted and managed using the principles of thermodynamics. Developers will eventually treat AI software stability as a problem of physics rather than just a problem of code.
From the abstract
We introduce κ-Desktop, an adaptive desktop automation agent that exhibits information phase transitions analogous to thermodynamic phase transitions in matter. The system routes user instructions through three execution pathways-fast (rule-based), medium (template-based), and slow (LLM-based)-governed by a single parameter κ ∈ [0, 1], the Information Precision Index. We derive and validate a growth law κ(n) = 0.95(1-e-0.25n) + 0.05 with coefficient of determination R 2 = 1.0000, and identify tw