How to design a multi-agent system that skips the LLM // TRAIN BRAIN

How to design a multi-agent system that skips the LLM

Github repo → https://goo.gle/race-condition
Previous episode → https://goo.gle/marathonagent
A thousand AI agents run a marathon, and almost none of them ever call the LLM.
In this multi-agent system deep dive, Casey West breaks down the one architectural decision behind Race Condition: a 1000-agent system built on Google's Agent Development Kit (ADK).
The question every AI engineer is wrestling with: when do you let an LLM decide, and when do you just write the code in a multiagent system? We trace one decision end to end, planning a marathon route, then show how the same idea (skip the LLM where you don't need it) scales to a thousand agents running on deterministic code.
What you'll learn:
* When to use an LLM vs deterministic logic
* The before_model_callback trick, keep the agent, skip the model
* Why route planning is deterministic (NP-hard + the Spine & Sprout algorithm)
* How 1,000 autopilot runners make 0 LLM calls
* Where the tokens actually go (the AI decides, the code runs)
* Scaling 1,000 stateless sessions with Redis
Chapters
00:00 - Intro: 1,000 AI agents that don't call the LLM
00:41 - When should an agent use an LLM?
01:02 - [Demo] Planning a marathon route
01:59 - Why Google Maps can't route a marathon
05:08 - Why the LLM Is the wrong tool (NP-hard)
05:40 - The deterministic spine & sprout algorithm
06:58 - Using AI Studio to choose the algorithm
09:00 - The trick: Skip the LLM with a callback
12:26 - before_model_callback — the reveal
17:50 - Autopilot runners: 1,000 agents, 0 LLM calls
21:31 - How many tokens? Where they actually go
23:28 - The second cost: Session state & redis
29:05 - Wrap up
More resources:
Google Agent Development Kit (ADK) → https://goo.gle/3PItVzL
Google ADK Community (Redis session service) → https://goo.gle/4ugzmUw
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 → https://www.youtube.com/playlist?list=PLIivdWyY5sqKnJOvP89yF8t9mWuzMTcbM
? Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
#AIAgent #GoogleADK #Gemini #MultiAgentSystem #AgenticAI #GoogleCloud
Speakers: Casey West, Annie Wang
Products Mentioned: Google Agent Development Kit, Gemini API, Agent Runtime, Google Cloud Pub/Sub, AlloyDB, Agent2Agent Protocol

Google Cloud Tech

Helping you build what's next with secure infrastructure, developer tools, APIs, data analytics and machine learning....

How to design a multi-agent system that skips the LLM

What’s going on with Developers and AI?

5 tips for using Antigravity 2.0 on enterprise codebases, planning phase

Automate M365 to Google Workspace Migrations with ADK multi-agents

Converting Oracle data types to PostgreSQL

Automate SAP order-to-cash with multi-agent AI & Google Agent Development Kit (ADK)

Automate project intake with multi-agent AI using MCP, Google ADK, Cloud Run, and BigQuery

AlloyDB AI: Revolutionizing Hybrid Search for PostgreSQL

Build long-running agents with Google’s Agentic Stack | The Agent Factory

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

[Demo] Autonomous ML Reliability - Data Center Network

[Demo] High Resolution Network Telemetry: Data Center Network

How to connect AI agents directly to your enterprise data: Introducing the AlloyDB remote MCP server

Stop rogue AI: Enforce policy & cut costs in Agent Development Kit (ADK)

Oracle packages: From black box to PostgreSQL

Vibe coding an AI racing coach in 2 weeks

Tuning custom Gemma models for high speed computer vision

How Whering architects cost efficient multimodal AI apps

Summarize millions of rows with one line of SQL: AI.AGG

AlloyDB Lakehouse Federation: Unified access to BigQuery and Google Cloud Lakehouse

The ABCs of agent building

What's new in Go

Next gen agentic architecture: Hands on with Gemini 3.5 & ADK

From laptop to planet scale: Deploying enterprise grade AI agents

Building production ready full stack apps with AI

How @CodeWithHarry helps developers upskill with AI

Moving AI agents beyond “Hello World” to real production

Eradicating friction for client side AI agents with Firebase

Mastering AI stacks for software engineers

Scaling multi-modal AI to 7 million users

Build an AI agent app with Go ADK, Cloud Run, and Flutter

Build full-stack apps with Google AI Studio, Cloud Run, and Cloud SQL

Develop and integrate AI agents with Google Workspace

Agent-first workflows from prompt to production

Stop over engineering your AI dev setup and just start

Dynamic Firebase skills: Architecting agent ready codebases

How Google Developer Experts vibecoded an AI racing coach with Gemini

Gemini Enterprise Agent Platform: Adding memory to AI agents

Firebase goes SQL: Inside the new SQL Connect (PostgreSQL)

Full-stack Dart is here: Top 5 Flutter highlights from Cloud Next '26

Stop building slow AI: Optimizing multi-agent systems for production

The future of Cloud AI: Mastering MCP servers, Gemini, and agentic workflows

Real time fraud detection with AlloyDB AI

Building conversational applications with Bigtable and ADK

Your AI agent still has no memory? Fix it with these 3 patterns

From systems of intelligence to systems of action: Yasmeen Ahmad on the agentic data cloud

Google Cloud Next '26 Developer Keynote recap

Automatic Read Scaling with AlloyDB Transparent Query Forwarding (TQF)

No design skills? Use this free AI powered design tool from Google | The Agent Factory

Your AI agent is forgetful. Here’s how to give it a brain.

Build cost effective, scalable apps using Firestore with MongoDB compatibility

How to scale Gen AI to billions of rows in BigQuery at a fraction of the cost

How does Bigtable work?

Welcome to Bigtable

Why 99% of AI agents fail in production (and how to fix it) | The Agent Factory

What is Gemini Enterprise Agent Platform?

Gemma 4 production stack: Model Armor, ADK Agents, Tracing

Self host Gemma 4: Deploy LLMs on Cloud Run GPUs

How to scale to 60 million requests (with ZERO ops team) using Cloud Run

AI agent long-term memory with memory bank

Mark your calendars for the AI Agent Clinic on April 15th!

Join our Googlers live on April 15th!

Orchestrating ML/AI workloads with TPUs on GKE

BigQuery Graph in 5'

How to add persistent memory to your AI agent

What are domain specific language models?

Cloud Run Networking explained (Updated!)

Fundamentals of ADK - Learning Path

Looker and AlloyDB: The ultimate stack for near real time operational business intelligence

How to authenticate Google Cloud Client Libraries

Reinforcement learning on TPU demo

Private routing to Google with Network Connectivity Center

Running a multi-agent AI architecture

How to get started with Google Cloud Client Libraries

BigQuery Migration Service: Validation and optimization

Build an AI agent with Gemini CLI and Agent Development Kit

Mainframe Connector demo series

Accelerate with AI debrief

Quickstart: Conversational Analytics with GCP Billing and Looker

Scaling your AI agent architecture with Cloud Run