Google Scion: A Hypervisor for AI Agents
Google just open-sourced the thing everyone has been building in-house but nobody has standardized -- a way to run multiple AI agents in parallel without them stepping on each other.
Scion is a multi-agent orchestration testbed that treats AI agents like virtual machines. Each agent -- whether it is Claude Code, Codex, Gemini CLI, or any containerized tool -- gets its own container, its own git worktree, and its own credentials. They can work on different parts of your project simultaneously, completely isolated from each other. Think Docker for agents, or more precisely, a hypervisor that manages agent lifecycles.
The architecture is deliberately minimal. Rather than enforcing rigid coordination patterns, Scion gives agents a CLI tool and lets the models themselves decide how to coordinate through natural language. There is a local mode that runs standalone, and a distributed Hub mode for spreading agents across remote VMs or Kubernetes clusters. It even supports tmux for attaching and detaching from running agent sessions.
At 485 stars and still labeled experimental (local mode is stable, Hub is about 80% there, Kubernetes is early), this is clearly Google Cloud Platform defining the operating system layer for multi-agent workflows. The Apache 2.0 license means it is meant to be adopted, not just observed.
The timing matters. Every team running more than one coding agent is solving this problem with duct tape -- tmux sessions, manual git branch management, hoping agents do not conflict. Scion says that is an infrastructure problem, not a workflow problem, and infrastructure should be shared.
https://github.com/GoogleCloudPlatform/scion
← Back to all articles
Scion is a multi-agent orchestration testbed that treats AI agents like virtual machines. Each agent -- whether it is Claude Code, Codex, Gemini CLI, or any containerized tool -- gets its own container, its own git worktree, and its own credentials. They can work on different parts of your project simultaneously, completely isolated from each other. Think Docker for agents, or more precisely, a hypervisor that manages agent lifecycles.
The architecture is deliberately minimal. Rather than enforcing rigid coordination patterns, Scion gives agents a CLI tool and lets the models themselves decide how to coordinate through natural language. There is a local mode that runs standalone, and a distributed Hub mode for spreading agents across remote VMs or Kubernetes clusters. It even supports tmux for attaching and detaching from running agent sessions.
At 485 stars and still labeled experimental (local mode is stable, Hub is about 80% there, Kubernetes is early), this is clearly Google Cloud Platform defining the operating system layer for multi-agent workflows. The Apache 2.0 license means it is meant to be adopted, not just observed.
The timing matters. Every team running more than one coding agent is solving this problem with duct tape -- tmux sessions, manual git branch management, hoping agents do not conflict. Scion says that is an infrastructure problem, not a workflow problem, and infrastructure should be shared.
https://github.com/GoogleCloudPlatform/scion
Comments