Adversarial stress-testing of long-horizon multi-agent systems