Every time my AI agents broke in production, my instinct was to reach for a bigger model and it almost never worked. A method I patented years ago in chaotic Indian traffic (a trapezoid bounding box instead of a rectangle) taught me why: the bottleneck is almost always how you represent the problem, not the size of the model. Six representation-first lessons on context, occlusion, evals, and the gap between a cool demo and an agent you'd trust to take real action for anyone shipping agentic A...