How we built 1,000 AI agents that run a marathon

Google Cloud Tech
AI summary

Casey West demonstrates how Google built a production-grade multi-agent system running 1,000 simultaneous ADK agents for a marathon visualization at Google Cloud Next 2026. The video covers the 4-layer polyglot architecture, switching from HTTP-based A2A to Redis pub/sub for dispatch, agent card discovery, and production patterns like resilient_model with retries, jitter, and exponential backoff. Developers building multi-agent systems will learn architectural decisions and trade-offs for scaling agent communications.