What Management Styles Teach About Agent Orchestration

Jeff ran four AI agents through an IKEA bookshelf assembly simulation using a 20B open-weights model, testing command-and-control, Taylorism, outcome-focused, and fully autonomous orchestration patterns. The results map neatly onto real engineering management. Outcome-focused coordination - define what done looks like, set WIP limits, stay out of the way - finished 5x faster than micromanaged command-and-control and 3x faster than rigid task decomposition. One agent spontaneously expanded its role to cover a bottleneck, emergent behavior that nobody programmed. Full autonomy without constraints produced chaos and duplicate work. The experiment lands at a moment when multi-agent frameworks from LangChain, CrewAI, and AutoGen are all converging on similar orchestration questions, and most default to the wrong management style.