Most agentic AI talks stop at orchestration. This one goes further.

This session demonstrates a production-grade multi-agent system built on Google ADK and Vertex AI Agent Engine — with Model Armor deployed as a real-time AI firewall at the inference layer.

Using a live image generation pipeline, we cover hierarchical agent design (LoopAgent, SequentialAgent), A2A cross-vendor interoperability, and MCP tool integration. Then we fire up the firewall.

The headline demo shows Model Armor intercepting threats before they reach the LLM: PII (credit cards, SSNs) blocked cold, harmful content stopped, every decision logged with reason codes for compliance audit.

Bonus: one env var swaps Gemini for Claude — zero agent code changes. Model-agnostic by design.

Attendees leave with a reference architecture, Google Codelab link, and a security pattern that makes AI governance a first-class design principle.