Active Inference Skill

"Action is the process of changing the world to match your prediction."

1. The Concept (Free Energy Principle)

Standard agents are goal-directed (maximize reward); Active Inference agents are surprise-minimizing (minimize prediction error).

• Goal: not just to "win" but to understand and control.
• Surprise: the gap between expectation and observation.

Implementation note (conceptual framing, not implemented CLI math): the production gabbe/brain.py uses epsilon-greedy gene selection plus a monotonic success_rate reward (effectively a bandit over prompt variants); the free-energy / active-inference framing here is conceptual, not implemented math. Use it to shape agent instructions, not as a literal solver — and never claim it "guarantees correctness." See guides/ai/self-evolving-skills.md for what actually runs.

2. The Feedback Loop

1. Predict: "Running go test will output PASS."
2. Act/Sense: run the command, read output.
3. Compare: compute prediction error — result "FAIL" → Surprise!

3. Solving the Error

Two ways to minimize surprise:

1. Perceptual Inference (change mind): "My model was wrong" → update docs/mental model.
2. Active Inference (change world): "The code is wrong" → edit it so the test passes.