Banger paper for agent builders. Multi-agent systems often underdeliver. The problem isn't how the agents themselves are built. It's how they're organized. They are mostly built with fixed chains, trees, and graphs that can't adapt as tasks evolve. But what if the system could learn its own coordination patterns? This new research introduces Puppeteer, a framework that learns to orchestrate agents dynamically rather than relying on handcrafted topologies. Instead of pre-defining collaboration structures, an orchestrator selects which agent speaks next based on the evolving conversation state. The policy is trained with REINFORCE, optimizing directly for task success. Rather than searching over complex graph topologies, they serialize everything into sequential agent selections. This reframing sidesteps combinatorial complexity. What emerges is surprising: compact cyclic patterns develop naturally. Not sprawling graphs, but tight loops where 2-3 agents handle most of the work. The remarkable part is that the system discovers efficiency on its own. Results: - On GSM-Hard math problems: 70% accuracy (up from 13.5% for the base model alone). - On MMLU-Pro: 83% (vs 76% baseline). - On SRDD software development: 76.4% (vs 60.6% baseline). These gains come with reduced token consumption. The paper shows that token costs consistently decrease throughout training while performance improves. They also prove the agent selection process satisfies Markov properties, meaning the current state alone determines the optimal next agent. No need to track full history. Why it matters for AI devs: learned simplicity beats engineered complexity. A trained router with a handful of specialized agents can outperform elaborate handcrafted workflows while cutting computational overhead.