Voting Protocols Boost Coordination in Multi-Agent AI Tutoring

Agentic tutoring systems face a fundamental coordination problem: multiple specialized agents may each propose reasonable but different interventions, yet only one response can be delivered to the learner. Eric Qiu and Joyce Gill address this by studying voting protocols as coordination mechanisms for four role-constrained pedagogical agents responsible for scaffolding, misconception handling, motivation, and metacognition. They compare four voting protocols—simple, ranked, cumulative, and approval—across two simulated tutoring environments using SciQ and HumanEval benchmarks. Rather than using voting as a mere aggregation step, the authors analyze how collective decision rules shape coordination under partial pedagogical conflict. Over 1,200 simulated interactions, they find that both agent deliberation and the type of voting protocol frequently change which response ultimately wins, demonstrating that both factors meaningfully shape collective decisions.

Different voting rules produce distinct coordination behaviors among the role-specialized agents, and even brief tutoring turns show measurable learning gains in simulated students. The study isolates how protocol choice is associated with distinct coordination patterns, offering practical insights for designing multi-agent tutoring systems that can effectively combine specialized expertise. By shifting from simple majority voting to more nuanced protocols like cumulative or approval voting, the system can better balance competing pedagogical priorities. This research, accepted at the ICML 2026 Workshop on AI4Good, suggests that thoughtful design of agent coordination mechanisms can directly improve learning outcomes without requiring more complex models.