Google has open‑sourced Scion, an experimental testbed that treats multi‑agent systems like a hypervisor. The platform spins each agent—Claude Code, Gemini CLI, Codex and others—into its own container, git worktree and credential set, letting them run locally, on remote VMs, or inside Kubernetes clusters. Isolation is the default, so agents can pursue independent goals for large‑scale projects without colliding.

Scion’s orchestration model lets developers define a dynamic task graph where agents execute in parallel, handling coding, auditing or testing steps as separate lifecycles. Long‑lived specialists coexist with throwaway workers that disappear after a single job. The system enforces guardrails and resource allocation through container policies and network rules rather than embedding restrictive scripts, effectively running agents in a “--yolo” mode while keeping the environment safe.

To illustrate the workflow, Google ships a demo game called Relics of the Athenaeum, where multiple agents collaborate to solve computational puzzles. Harnesses called “adapters” manage each agent’s lifecycle, authentication and configuration, supporting Gemini, Claude Code and partially Codex. Developers can pick Docker, Podman, Apple containers or Kubernetes via named profiles and versioned, giving them flexibility to embed Scion into existing CI pipelines.