What is that bump in utility in practical terms? You can point to a benchmark improvement but that's no indication the agent swarm is not reducing to "giving an llm an arbitrary amount of random guesses".