Google has open-sourced Scion, an experimental multi-agent orchestration testbed designed to run concurrent AI agents in isolated containers across local machines and remote clusters. Each agent gets its own container, Git worktree, and credentials, preventing file and credential collisions when multiple agents work on the same project. Rather than hard-coding manager/worker roles, Scion uses an emergent coordination model where agents dynamically learn a CLI tool and self-organize task division. The architecture supports Claude Code, Gemini CLI, and OpenAI Codex as agent runtimes, with a layered configuration system that scales from local Docker to Kubernetes without rewriting orchestration logic. Still in alpha, it's positioned as a testbed for developers exploring parallel agent execution patterns.
Table of contents
What Makes Scion DifferentA “Less is More” PhilosophyArchitecture at a GlanceWhy it Matters NowSort: