Summary
A survey of research applying the Free Energy Principle (FEP) and Active Inference (AIF) to software systems from 2024–2026 reveals two dominant streams and a critical gap. The infrastructure stream uses multi-agent active inference for fault diagnosis, resilience, and service management in distributed edge/cloud environments. The code entropy stream applies information-theoretic measures (Boltzmann, Tsallis, structural/textual entropy) to predict software aging, bugs, and trustworthiness decay. A single attempted bridge—Alexandria Code Intelligence—uses neuroscience terminology but remains static analysis, not a genuine variational free-energy minimizer. No project yet implements the full FEP loop (generative model → prediction → prediction error → model update → action) with a codebase as the environment. This leaves the territory of self‑modeling code evolution unoccupied.
Key Points
- Infrastructure projects (AURORA, PAIR‑Agent, Distributed Intelligence, Active Inference Digital Twins) model distributed services with active‑inference agents, achieving >90% SLO fulfillment in one case.
- Code entropy projects (SGT, Boltzmann entropy for software aging, entropy‑based bug prediction, information theory in software evolution) analyze code structure statistically but do not generate predictions or actions.
- Alexandria Code Intelligence is the only attempt to bridge the two streams, but it is static analysis dressed in neuroscience metaphor—not a genuine active‑inference system.
- Gap: A system that treats its own codebase as an environment, maintains a generative model, and updates it through prediction‑error minimization does not exist.
Concepts
- Free Energy Principle (FEP): A theoretical framework from neuroscience stating that adaptive systems minimize variational free energy (surprise) by updating internal models or acting on the environment.
- Active Inference (AIF): A corollary of FEP where agents choose actions to minimize expected free energy, balancing exploration and goal‑seeking.