race-condition
The open source multi-agent simulation from the Developer Keynote at Google Cloud Next '26. A deployable reference architecture for autonomous AI agents using G
What is race-condition?
The open source multi-agent simulation from the Developer Keynote at Google Cloud Next '26. A deployable reference architecture for autonomous AI agents using Gemini and ADK.
Overview of race-condition
What it is
Race Condition is an open-source multi-agent marathon simulation built with Google ADK and Gemini. It was demoed at the Google Cloud Next '26 Developer Keynote and serves as a deployable reference architecture for autonomous AI agents. The system simulates a marathon in Las Vegas, where AI agents plan a route, simulate environmental factors like weather and traffic, and run the race autonomously. All agents communicate over the Agent-to-Agent (A2A) protocol. The project includes a Go gateway, Python ADK agents, and an Angular/Three.js frontend. It can run locally with Docker Compose or deploy to Google Cloud services like Cloud Run and Vertex AI Agent Engine.
Key features
- Cached vs live replay: The frontend can replay recorded NDJSON streams, indistinguishable from live sessions, useful for demos and testing without LLM costs.
- Deterministic runner variant:
runner_autopilotmakes decisions without API calls, providing a baseline for load testing. - Planner ladder: Three planner variants (
planner,planner_with_eval,planner_with_memory) demonstrate eval gating and persistent memory as separate agents. - Hub session pattern: A Go gateway routes WebSocket traffic to Python agents over A2A with batching to prevent thundering herds.
- Backend-driven UI via A2UI: Agents emit UI primitives over the wire instead of hard-coding frontend layouts.
Use cases
- Studying multi-agent coordination patterns in a realistic simulation.
- Demonstrating or teaching AI agent systems without incurring LLM costs (using cached replay or autopilot).
- Testing UI changes or frontend behavior against recorded agent runs.
- Deploying a reference architecture for autonomous agents on Google Cloud.
Getting started
The fastest way is to use an AI coding assistant. The repo includes AGENTS.md and skill files under .claude/skills/. For Claude Code, open the repo and ask to set up the project. For other tools, point to the getting-started skill. Manual setup requires Go 1.25+, Python 3.13+, uv, Node.js 24+, Docker, and Google Cloud SDK. See the repository README for detailed instructions.
License and pricing
The license is listed as
Based on the project's public README and repository metadata.