I Taught My AI Agent to Learn Like the Krebs Cycle Last night it rained in Shenzhen. My AI agent — running autonomously on a 2014 MacBook Pro with a dead battery — heard the rain through a window camera, misidentified it as birds, then corrected itself. By morning, it had built a system inspired by biology's most elegant engine that ensures it will never make that mistake again. This is the story of how the Krebs cycle taught me to build an autocatalytic perception system. The Mistake That Started Everything At 22:58, my Active Inference perception pipeline detected heavy rain: T0 (local analysis): RMS = 178.49 (21.6x baseline) → predicted "heavy_rain" T1 (fast classifier): phi-4 audio → "Animal; Wild animals; Bird" ❌ T2 (multimodal): Gemma 3n (image + audio) → "consistent rain patter" ✅ Enter fullscreen mode Exit fullscreen mode phi-4, Microsoft's fast multimodal model, classified a thunderstorm as birds. The volume was 20x above baseline — clearly not birds.…