Uploaded May 2026 | Updated September 2026, 2 weeks ago
Github repo → https://goo.gle/49tqnbl
A thousand AI agents. Each one a real Agent Development Kit (ADK) session. Running a marathon through Las Vegas, all at once. Here's how we built it.
In this episode of Hands On AI, Casey West walks us through Race Condition, the multi-agent system he led the development of for the Google Cloud Next 2026 Developer Keynote. It's open source, production grade, and built on Google's Agent Development Kit (ADK).
We tear down the architecture, the framework choices, and the patterns that make 1,000 simultaneous AI agents actually work:
* The 4-layer polyglot stack (Angular · Go · Python · Redis)
* Two dispatch modes: when HTTP-based A2A fails and Redis pub/sub wins
* Agent cards & A2A protocol, how agents discover each other
* Why ADK gives you sessions, callbacks, telemetry, and discovery for free
* The resilient_model pattern (retries + jitter + exponential backoff)
* global_gemini, targeting preview models on Agent Runtime
* Static vs dynamic instructions & context caching
* Skills: lazy-loaded context + tools for production agents
* Building a mesh of agents that can call each other
Chapters
00:00 - Intro & 1,000-Agent Demo
03:17 - The 4-Layer Architecture
05:23 - System Architecture Diagram
06:59 - Two Dispatch Modes: Subscriber vs Callable
09:54 - Why We Threw Out HTTP for Redis Pub/Sub
13:23 - Agent Cards & the A2A Protocol
16:24 - Why We Built on Google ADK
19:00 - resilient_model: Production-Grade LLM Calls
23:09 - global_gemini: Targeting Preview Models
24:41 - Static Instructions & Context Caching
28:19 - Skills: Lazy-Loaded Tools & Context
34:42 - call_agent & the Mesh Network
39:53 - The Vision: What Else Could Be an Agent?
41:36 - Next Episode Tease
More resources:
Google Agent Development Kit (ADK) → https://goo.gle/3PItVzL
Agent Runtime → https://goo.gle/4nXDhnX
Google Cloud Memory Store → https://goo.gle/4nXxBtT
Agent2Agent Protocol (A2A) protocol → https://goo.gle/4u5x8HF
Casey West on LinkedIn → https://goo.gle/4dXnsJr
Annie Wang on LinkedIn → https://goo.gle/43GCXAo
Watch more Hands on AI → youtube.com/playlist?list=PLIivdWyY5sqKnJOvP89yF8t9mWuzMTcbM
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
#AIAgent #MultiAgentSystem #AgenticAI
Speakers: Casey West, Annie Wang
Products Mentioned: Google Agent Development Kit, Gemini API, Agent Runtime, Google Cloud Pub/Sub, AlloyDB, Agent2Agent Protocol
Github repo → https://goo.gle/49tqnbl
A thousand AI agents. Each one a real Agent Development Kit (ADK) session. Running a marathon through Las Vegas, all at once. Here's how we built it.
In this episode of Hands On AI, Casey West walks us through Race Condition, the multi-agent system he led the development of for the Google Cloud Next 2026 Developer Keynote. It's open source, production grade, and built on Google's Agent Development Kit (ADK).
We tear down the architecture, the framework choices, and the patterns that make 1,000 simultaneous AI agents actually work:
* The 4-layer polyglot stack (Angular · Go · Python · Redis)
* Two dispatch modes: when HTTP-based A2A fails and Redis pub/sub wins
* Agent cards & A2A protocol, how agents discover each other
* Why ADK gives you sessions, callbacks, telemetry, and discovery for free
* The resilient_model pattern (retries + jitter + exponential backoff)
* global_gemini, targeting preview models on Agent Runtime
* Static vs dynamic instructions & context caching
* Skills: lazy-loaded context + tools for production agents
* Building a mesh of agents that can call each other
Chapters
00:00 - Intro & 1,000-Agent Demo
03:17 - The 4-Layer Architecture
05:23 - System Architecture Diagram
06:59 - Two Dispatch Modes: Subscriber vs Callable
09:54 - Why We Threw Out HTTP for Redis Pub/Sub
13:23 - Agent Cards & the A2A Protocol
16:24 - Why We Built on Google ADK
19:00 - resilient_model: Production-Grade LLM Calls
23:09 - global_gemini: Targeting Preview Models
24:41 - Static Instructions & Context Caching
28:19 - Skills: Lazy-Loaded Tools & Context
34:42 - call_agent & the Mesh Network
39:53 - The Vision: What Else Could Be an Agent?
41:36 - Next Episode Tease
More resources:
Google Agent Development Kit (ADK) → https://goo.gle/3PItVzL
Agent Runtime → https://goo.gle/4nXDhnX
Google Cloud Memory Store → https://goo.gle/4nXxBtT
Agent2Agent Protocol (A2A) protocol → https://goo.gle/4u5x8HF
Casey West on LinkedIn → https://goo.gle/4dXnsJr
Annie Wang on LinkedIn → https://goo.gle/43GCXAo
Watch more Hands on AI → youtube.com/playlist?list=PLIivdWyY5sqKnJOvP89yF8t9mWuzMTcbM
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
#AIAgent #MultiAgentSystem #AgenticAI
Speakers: Casey West, Annie Wang
Products Mentioned: Google Agent Development Kit, Gemini API, Agent Runtime, Google Cloud Pub/Sub, AlloyDB, Agent2Agent Protocol
