Unifying AI experience for customers: Building multi-agent systems with Google ADK
summary
Built a supervisor “super-agent” that routes to specialist sub-agents using Google ADK and the A2A protocol, with a live POC demo showing multilingual responses, guardrail rejections, and per-agent auth. Routing uses a GEM 3.0-Z 32B supervisor for intent selection, sub-agents use smaller models, and context is summarized and injected at runtime with a 256-token thinking budget. Observability surfaces 10–15s latencies, token counts and tool calls via Lenfuse so failures and routing errors can be traced and evaluated.
AI-written summary · about
chapters (8)
- 0:00Multi-agent systems intro
- 4:33Agents and AI silos problem
- 8:25Fragmentation with supervisor agents
- 10:56Challenges in multi-agent systems
- 17:38Architecture and observability
- 23:42Why Google ADK
- 26:39Live demo
- 35:42Q&A on latency
related talks
54mCreate Multi-Agent Systems with ADK
01h 09mMulti Agent Architecture using ADK
01h 03mSupercharge your AI agents: Inside the new ADK integrations ecosystem
31mBuilding, Deploying and Managing AI Agents at Scale
39mA2A in 45 Minutes: Building Interoperable AI Agents
31mA2A in 30 Minutes: Building Communicating AI Agents
01h 00mBuilding a Polyglot Multi-Agent AI Architecture
43mA2A in 50 Minutes: Building Interoperable AI Agents