A2A Multi-Agent Demo — Agent Engine × Apigee × Cloud Run
A replicable Google Cloud reference implementation of an enterprise-governed multi-agent system: a supervisor agent delegates to three specialist agents over the A2A protocol, with every model, MCP tool and agent call routed through Apigee for auth, quotas and token analytics.
This project is a complete, deployable reference implementation showing how to govern a multi-agent AI system on Google Cloud rather than a toy demo: a supervisor agent running on Vertex AI's Gemini Enterprise Agent Platform delegates to three domain-specialist agents over the Agent2Agent protocol, and Apigee sits in front of every data path - calls to both Gemini and Claude models, MCP tool calls, the agents themselves, and the e-commerce backend the tools query - with every hop correlated by a single trace ID in one observability pane.
Deploying it provisions real security boundaries (Identity-Aware Proxy, IAM-only services, private networking), per-user API keys and token quotas bound to signed-in identity, a trace explorer that replays any transaction across the full three-project topology with a latency waterfall, and optional Apigee and Cloud Monitoring analytics dashboards for token usage per model, agent and user. Provisioning tooling rebuilds the entire stack from three empty Google Cloud projects, driven from one environment file and three declarative manifests using a consistent check/apply command grammar.
It is built with Google's Agent Development Kit and targets Cloud Run and Vertex AI Agent Engine.