Minimal Resource-Constrained Agent¶

Status: executable toy controller. Its variables are engineering proxies, not measured thermodynamic state variables and not a definition of intelligence.

Purpose¶

The code in core/constraints.py and core/minimal_agent.py compares a controller that trades task reward against a resource budget with one that pursues the task reward more aggressively. It asks a small question: can an explicit cost and stability term alter the controller's trajectory in the stipulated environment?

Variables¶

energy is a numeric action budget. The values assigned to actions are chosen by the model; they are not joules unless an external calibration supplies units.
entropy is an accumulated damage or disorder proxy. Adding one per step and ten for an aggressive action is a simulation rule, not thermodynamic entropy.
stability_score is the reciprocal of that proxy. It reports the chosen state variable and does not establish homeostasis.
task_success is the reward supplied by the environment.

The combined objective is a multi-objective control rule. Referring to it as free energy is only an analogy; the script does not derive a variational free-energy objective or implement active inference.

What a Run Shows¶

Under the selected parameters, the constrained controller can preserve more of its budget or stability proxy while accepting less immediate task reward. That is an ordinary trade-off produced by the reward and transition equations.

It does not show:

that intelligence requires these variables;
that the coefficients are physically correct;
that one scalar captures ecological or biological viability;
that a constrained controller is aligned;
that the policy generalizes outside the toy environment.

Useful Follow-Ups¶

Sweep the cost coefficients and initial states, match total action opportunity, add delayed and noisy measurements, compare hard budgets with soft penalties, and introduce an adversary that can alter the proxy. Report a Pareto frontier rather than declaring one weighting optimal.

The module is valuable as a minimal test harness for constraint design. Its strongest defensible claim is that explicitly represented costs can change behavior in the model.