Information Thermodynamics of Agents: The Work Capacity of Channels with Memory

Thermodynamic computers are stochastic physical devices designed to perform calculations at the thermal energy scale. Their operation is constrained by the equations of stochastic thermodynamics, among which are a set of bounds, known as speed limits, that relate a thermodynamic computer's run time to its computational progress and the heat it dissipates. Using the Wasserstein speed limit we assess the thermodynamic efficiency of a simulation model of a thermodynamic computer trained to perform a standard machine-learning classification task. On this task the thermodynamic computer is as capable as a simple multilayer perceptron. We show that different inference protocols allow the computer to operate within 40% of the thermodynamic limit of efficiency without loss of accuracy, or to perform inference increasingly rapidly at fixed accuracy and thermodynamic efficiency. These results indicate that a thermodynamic computer designed for a particular task retains considerable freedom in its thermodynamic operation.

Adaptive behavior under partial observability depends on internal organization that carries information beyond the current observation. Drawing on Barrett and Miller's account of categorization as predictive, compressive, functionally organized, and allostatically constrained, we test whether recurrent and spiking agents develop internal states with corresponding computational properties. Agents operate in an energy-constrained foraging task requiring resource acquisition, threat avoidance, contact-dependent consumption, and regulation of an internal energy variable. In a frozen benchmark, learned agents outperform random and heuristic baselines; the trace-augmented recurrent policy is strongest overall, while spiking variants show stress-specific differences. Early internal dynamics predict later full-safe-efficient success above permutation baseline, reaching a maximum ROC-AUC of 0.802. Reduced PCA subspaces retain behaviorally relevant information. Feature-family controls show that predictive signal is distributed across trace, policy-head, internal-dynamics, observation, and allostatic variables, and low-energy state remains strongly decodable after explicit energy-related features are removed. Evaluation-time perturbations to temporal state, sensory information, operating conditions, and allostatic mechanisms alter behavior and/or internal prediction. Seed-balanced event probes show weaker but measurable information about future contact, successful consumption, and threat events, alongside strong low-energy decoding. We interpret this pattern as a computational analogue of predictive allostatic organization: distributed control regimes that are predictive, energy-sensitive, action-relevant, and partly causally involved, without claiming biological validation or discrete symbolic categories.

We develop a thermodynamic theory of algorithmic catalysis within the watts per intelligence framework, identifying reusable computational structures that reduce irreversible operations for a task class while satisfying bounded restoration and structural selectivity constraints. We prove that any class specific speed-up is upper-bounded by the algorithmic mutual information between the substrate and the class descriptor, and that encoding this information incurs a minimum thermodynamic cost via Landauer erasure. Combining these results yields a coupling theorem that lower-bounds the deployment horizon required for an algorithmic catalyst to be energetically favourable. The framework is illustrated on an affine SAT class and situates contemporary learned systems within an information thermodynamic constraint on intelligent computation.