CyberAGENTS: How Structured Autonomy Shapes Multi-Agent Behavior in Gamified Cybersecurity Learning

Multi-agent systems are increasingly used in interactive learning, yet evaluations often emphasize aggregate outcomes rather than what agents and humans do during runtime. We study how schema- and ontology-based constraints shape observable behavior in CyberAgents, a multi-agent system for gamified cybersecurity learning. Our analysis considers three levels of behavior: how role-specialized agents act and transition during a run, how learners perceive and respond to agent guidance, and how interaction patterns change when structural constraints are removed. We analyze 208 logged interaction trajectories from full (n=111) and ablated (n=97) deployments using automated rubric-based behavioral coding, transition analysis, and post-interaction feedback from 24 undergraduate students. In the full system, all four agent roles appeared in 70.3% of trajectories, compared with 47.4% in the ablation. Automated coding also found higher role fidelity, learner-intent alignment, interaction coherence, and cybersecurity grounding in the full configuration. Behavioral failures such as overly generic responses, ignored learner corrections, and role confusion occurred substantially more often in ablated trajectories. Learners rated scenario authenticity and trust in agent feedback positively, while identifying feedback length, interaction transitions, and challenge calibration as recurring difficulties. These results show how explicit structural constraints are associated with observable differences in multi-agent sequencing, behavioral failures, and human–agent interaction quality, while also illustrating the value and limitations of automated behavioral analysis.