Google's Scion: A Testbed for Orchestrating Concurrent Agents

Scion is an experimental orchestration testbed developed by Google, designed to manage concurrent agents running in containers across both local and remote compute environments. It acts as a 'hypervisor for agents,' integrating components like agent memory, chatrooms, and task management. Scion orchestrates 'deep agents' such as Claude Code, Gemini CLI, and Codex, each operating in isolated containers with their own credentials and git worktrees. This setup allows agents to work independently on different project aspects without interference.

Scion's architecture supports dynamic task graphs that evolve and execute in parallel, catering to varied goals like coding, auditing, and testing. It emphasizes isolation over constraints, allowing agents to operate freely within defined boundaries. This approach is exemplified by running agents in '--yolo mode,' ensuring safety through containerization and network policies.

The platform supports multiple agents via harnesses that manage lifecycle, authentication, and configuration. While agents like Gemini and Claude Code are fully supported, Codex and OpenCode have partial support. Developers can choose from various containerization runtimes, including Docker, Podman, and Kubernetes.

To effectively use Scion, developers must understand its unique terminology, such as 'grove' for projects and 'hub' for orchestration control planes. Google has demonstrated Scion's capabilities through the game 'Relics of the Athenaeum,' where agents collaborate to solve puzzles, showcasing dynamic agent spawning and collaboration through shared workspaces and communication channels.