Multi-Agent AI for Production Security Operations: A2A and MCP Architecture in 5G Core Networks

The architecture is designed for a production 5G core network, which generates more security telemetry per hour than a typical enterprise SOC receives in a week. It replaces two ineffective mainstream patterns: a single LLM struggling with context window limits, indeterminism, and change management; and GenAI integrated into legacy SIEM/SOAR systems that still require manual rule writing. The proposed multi-agent system features narrow agents with single-page contracts, coordination via the open A2A protocol (open-sourced by Google and donated to the Linux Foundation), environment integration via MCP (donated by Anthropic), a privileged reviewer agent that uses OPA and Kyverno for deterministic policy enforcement, and human-in-the-loop as a first-class output. Before any LLM invocation, an Isolation Forest model pre-filters telemetry, sending only anomalous samples to the LLM-driven agents. The reviewer agent uses OPA for business safety invariants (blast radius, confidence, reversibility) and Kyverno for platform security invariants (trusted registries, network policies). Measured results include a 40% improvement in MTTD and MTTR, over 80 auto-generated detection rules, and a 12x reduction in manual effort for new rules. The system can monitor 10-20 5G network functions concurrently. Limitations apply when rule count is low, detection logic must be deterministic, or engineering functions do not cover reviewer security policies. Key lessons include building the reviewer agent first, adding anomaly pre-filtering from day one, avoiding clever prompts, and publishing reference patterns early.