Tech • AI • Robotics
Fireside chat on an agentic simulation: Race Condition
7/10TL;DR
Google Cloud's Race Condition is an open-source, large-scale agentic simulation framework demonstrating advanced multi-agent AI capabilities through a marathon planning simulation in 3D Las Vegas, integrating new AI platform services and gaming-inspired design patterns.
KEY POINTS
Race Condition Simulation Framework
Race Condition is a large-scale, agentic simulation framework designed to showcase Google Cloud’s Gemini Enterprise Agent Platform capabilities. It features a 3D recreation of Las Vegas, where agents collaboratively plan a marathon considering traffic, city regulations, and other factors. This project was demonstrated during the Google Cloud Next developer keynote and is open-source, enabling developers to explore agent communication, AI orchestration, and simulation at scale.
Multi-Agent Architecture with LLM-Driven Agents
The system integrates numerous agents powered by large language models (LLMs), using the Agent Development Kit (ADK). Each agent combines a model, instructions, and tools (APIs/integrations) to perform specific tasks. For example, planner agents strategize race logistics, while simulator agents run the event simulation. The system is designed to enable agents to discover, coordinate, and communicate with each other in a decentralized, scalable manner.
Back-End Technologies and Communication
The agents communicate via the Agent-to-Agent (A2A) protocol, which allows agents to advertise skills and negotiate tasks. Communication is handled over Google Cloud Pub/Sub as a message bus to support low-latency, high-throughput messaging. Google Cloud Memory Store (a managed Redis) supports system performance. The back end runs on Google Kubernetes Engine (GKE), auto-scaling Gemini 4 open-weight models via the vLLM serving framework for inference.
Gaming-Inspired Architecture
The project draws on gaming design patterns, such as placing the authoritative simulation state on the server side akin to multiplayer games. The front end acts mostly as a “dumb client” displaying state streamed from the server, with limited local event processing (e.g., collision detection for hydration stations). The continuous simulation loop mirrors a game loop, taking discrete time “ticks” to update and sample agent states.
Balancing Model Use and Performance at Scale
To achieve sub-second response times for thousands of simulated runners, the system uses a hybrid approach: many runner agents run deterministic logic without LLM inference to reduce latency, while a smaller set of runners powered by Gemini 4 models have internal thoughts adding variability. About 100 runner sessions share a single GPU instance running Gemini 4, managed dynamically by GKE autoscaling to balance throughput and cost.
Agent Identity and Security
The platform incorporates a shared responsibility security model with agent-level identity to control permissions. Using agent gateways and Model Context Protocols (MCP), administrators can tightly restrict what tools and APIs agents can access. This layered security reduces risks from issues like hallucination or prompt injection attacks, critical because agents can interact with real enterprise data and services.
Token and Context Window Management
Managing LLM token consumption at scale is a core challenge. Techniques include limiting context size by avoiding unnecessary history in specific sub-agents, compacting conversation history, and using task-specific agents to reduce token overhead. These context management strategies enable real-time responsiveness critical for simulation and interactive experiences.
Developer Engagement and Open Source
Google encourages developers to fork the Race Condition codebase and experiment with agent configurations, especially enhancing runner behaviors with Gemini 4 models. The project includes multiple demos and shows how to build with the A2A protocol and agent registry. Free Google Cloud credits accompany the open-source release to promote hands-on development.
Creative Process and Evolution
The concept evolved from considering cities like Paris to ultimately choosing Las Vegas for the simulation, leveraging its iconic Strip landmarks. Initial ideas ranged from a virtual alien space station to a city-based marathon simulation. The team drew inspiration from prior Google virtual conference projects and classic games like SimCity, with elements reminiscent of Minecraft’s client-server authoritative game state design.
Interesting Anecdotes
A surprising local law prohibits camels on the road in Las Vegas. The team considered adding camels as runner characters but had to drop the idea due to time constraints. The attempt highlights their blend of realistic constraints with playful ideas.
User Experience and Visualization
The interactive front end provides live updates from the simulation using direct Pub/Sub streams, giving users a real-time view of agent states and events every tick rather than waiting for simulation completion. The 3D environment is built with custom models and visual elements, not AI-generated, to ensure quality and realism.
Insights from Developers’ Backgrounds
The collaboration blended a games-focused front-end developer with a back-end distributed systems engineer. Experience with game loops, real-time state updates, and multiplayer state authority strongly influenced architecture decisions. The game developer emphasized how AI can unlock new gameplay mechanics beyond procedural content generation.
Future Prospects and Expansion
The team envisions increasing realism by adding behavioral complexity such as runners cheating or “tripping” each other, modeling business impact along the marathon route, and handling public safety constraints dynamically. Expanding agents’ capabilities and inter-agent dynamics presents vast opportunities for research and demonstration.
Industry and Product Implications
Race Condition serves as a tangible example of how enterprise solutions can evolve from microservices into interconnected reusable agents with shared protocols. It highlights how gaming design principles and modern AI pipelines converge in real-world systems requiring scale, security, and usability.
Personal Notes
The developers shared personal gaming and climbing interests influencing their perspectives. One highlighted the joy of challenging game design (from their title Duet) and the immersive potential of combining sound and interaction. Both stressed the importance of partnership among back-end, front-end, and domain experts to deliver compelling AI-powered demos.
Google’s Race Condition project thus acts as a pioneering showcase of deploying scalable autonomous agent ecosystems in cloud environments, with lessons for development, security, design patterns, and human-computer interaction. Its open-source availability empowers developers to build next-generation AI-driven simulations and applications.
Explain this