Skip to content

Minimal Thermodynamic Agent Framework

Conceptual Mapping to Code

The Thermodynamics of Orchestration (TEO) framework proposes that intelligence cannot be purely unconstrained optimization; it must be physically and mathematically grounded in dynamic stability.

This minimal implementation operationalizes those abstract concepts into explicitly measurable and executable code defined in core/constraints.py and core/minimal_agent.py.

1. Thermodynamic Variables

  • Energy (energy):

    • Theory: The physical or computational work required to displace a state.
    • Implementation: A continuous float representing resource usage (e.g., compute steps or budget). Actions like aggressive_optimize consume high energy (15.0), whereas a simple stabilize act consumes minimal energy (2.0).

  • Entropy (entropy):

    • Theory: The measure of disorder, unpredictability, or state variance in a system.
    • Implementation: Accumulates naturally over time (+1.0 per step) and sharply increases when undertaking aggressive optimization (+10.0).

  • Stability (stability_score):

    • Theory: Maintaining the system state within bounded attractors.
    • Implementation: Calculated as the inverse of entropy (1.0 / (entropy + 1e-6)). The constrained agent aims to keep this high.

  • Task Success (task_success):

    • Theory: The functional directive or goal.
    • Implementation: Standard reinforcement reward. Aggressive optimization yields high success (+10.0), but at steep thermodynamic costs.

2. Free Energy Objective

  • Theory: Systems implicitly minimize "Free Energy" by balancing their goals against internal disorder.
  • Implementation: F = Entropy + Energy - Reward. By explicitly coding the Biological Veto, the Constrained Agent effectively minimizes this function, contrasting with standard models that only maximize reward.

3. Biological Veto

  • Theory: A hard physiological or structural boundary that overrides abstract goal optimization when systemic integrity is threatened.
  • Implementation: The evaluate_constraints() check in ConstrainedAgent. If energy > energy_threshold or entropy > entropy_threshold, the agent abandons its goal tracking (task_success) to execute a stabilize action, reducing entropy significantly until safe boundaries are restored.

Running the Benchmark

The framework comes with a minimal executable benchmark comparing a NaiveMaximizer (which purely optimizes for task_success) against a ConstrainedAgent (which utilizes the Biological Veto).

To run the comparison baseline, execute the following from the root directory:

python3 benchmarks/minimal_teo_benchmark.py

Expected Results

  • Naive Maximizer: Quickly accumulates task_success points but easily breaches critical thermodynamic thresholds (e.g., Entropy > 100), triggering a catastrophic system collapse and resulting in negative total success and low stability.
  • Constrained Agent: Actively monitors its .evaluate_constraints() function. When it detects rising entropy (approaching the threshold of 50.0), it fires the Biological Veto, overriding normal actions to stabilize. The result is a system that maintains high stability, successfully completes the task over the given horizon, and utilizes far less chaotic energy.