Governing Multi-Agent AI Systems in the Enterprise: Model Context Protocol and Agent-to-Agent Communication with MuleSoft Agent Fabric

The emergence of multi-agent AI architectures in enterprise software calls for clear protocols for inter-agent communication, governance frameworks, and observability tooling. This paper examines two protocols shaping that landscape: the Model Context Protocol (MCP) and the Agent-to-Agent (A2A) communication protocol, both implemented through MuleSoft Agent Fabric on Anypoint Platform. Drawing from first-hand participation in the TDX 2026 hands-on lab "Connect & Govern MCP and A2A with MuleSoft Agent Fabric," we present a technical analysis of the four-pillar Agent Fabric framework (Discover, Orchestrate, Govern, Observe), a concrete enterprise use case involving automated customer complaint resolution, and an empirical evaluation of two design claims. RQ1 (Parallelism): LLM-orchestrated parallel sub-agent execution reduced mean end-to-end latency by 26.5% over sequential execution (28,274 ms to 20,769 ms) and by 42.1% under a maximally-parallel configuration, across 30 trials per setting (Kruskal-Wallis H = 79.1, p < 0.0001; Cohen's d ≥ 5.97). RQ2 (PII Governance): The a-two-a-pii-detector v1.0.1 policy achieved 94.0% accuracy, 100% precision, 87.5% recall, and F1 = 0.933 across 50 test cases. Zero false positives were observed. Three false negatives arose from format-normalisation gaps (undelimited 9-digit strings and whitespace-padded delimiters). Gateway-blocked requests resolved in 64 ms versus 382 ms for passing requests, confirming that centralised enforcement imposes negligible overhead.