A Protocol-Governed Multi-Agent Tutor for Higher-Education Concept Explanation and Programming Support

Large language model tutors can explain concepts, generate quizzes, and assist with programming, but many current systems still combine planning, retrieval, tool use, and response generation inside a single agent. That design makes it difficult to control external actions, insert safety checks, and interpret failures. This paper presents a protocol-governed multi-agent tutoring prototype that uses Agent2Agent (A2A) coordination for inter-agent routing and the Model Context Protocol (MCP) for structured tool access. The prototype separates dialogue management, orchestration, retrieval, teaching, evaluation, code execution, safety review, and learner modeling into explicit roles with least-privilege tool scopes, consent-gated execution, schema validation, sandboxed code running, and structured audit logs. We report an internal pilot benchmark of 25 higher-education tutoring tasks spanning concept explanation/chat (n = 7), quiz generation/evaluation (n = 11), and programming support (n = 7). The system completed 22 of 25 tasks successfully (88.0%), with mean end-to-end latency of 2.00 s and 0.32 tool calls per task. Performance varied strongly by workflow: explanation and quiz tasks succeeded on all evaluated cases, whereas code-oriented tasks achieved 57.1% success and emerged as the main reliability bottleneck. The pilot recorded 14 safety checks and no observed policy violations, but it did not include a strong adversarial unsafe-tool stress test. The contribution of the paper is therefore not a claim of classroom learning gains; it is a transparent empirical systems account of how a protocol-first tutoring architecture