GoogleCloudPlatform / scion

# Scion
An open foundation for agent collaboration.
_sci·on /ˈsīən/ — a young shoot or twig, cut for grafting or rooting._
Scion is an open-source orchestration platform for teams of AI agents and the people working with them. Bring your choice of agent harnesses and models, give agents specialized roles, and let them delegate tasks, share findings, and work in parallel — on your laptop or across cloud infrastructure.
As agents take on longer, more complex projects, the challenge becomes how they work together: who owns a task, what context they share, when they ask for help, and how you follow their progress. Scion provides the running environments, communication, identity, and visibility to explore those questions in practice.
Run harnesses such as Claude Code, Antigravity, Codex, and OpenCode as independent processes, with per-agent configuration and credentials. Choose shared workspaces or separate Git worktrees and clones to suit the work. Start with a local CLI or the Workstation web UI, then move to a hosted Hub when your team needs shared access and remote execution.
Scion's philosophy is **less is more**. Agents learn the `scion` CLI through skills and help text; you describe responsibilities and collaboration patterns in instructions. Teams can grow, delegate, review, and regroup as a task unfolds. Use Scion as a practical testbed for emerging patterns in agent collaboration, with the freedom to choose your models, tools, and infrastructure. Read more in Philosophy.
## In Action
Coding harnesses give agents access to files, tools, and a working environment. Scion puts those capabilities to work across software development, research, testing, and creative projects. Internal experiments at Google have included software porting, market research, and product testing, with cloud operations and scientific research under exploration. The projects below show how collaboration patterns can be developed through real tasks and human feedback.
### Scion Films
In Scion Films, agent crews explored filmmaking, where success depends on human judgment as well as technical execution. Viewer feedback and agent retrospectives shaped the team's skills, tools, and collaboration across a series of pilots. An agent documentary crew then used those same practices to tell the story of the experiment.
### Relics of Athenaeum
Relics of Athenaeum is an "agent game" that demonstrates multi-agent orchestration defined entirely in markdown. A group of agents collaborate to solve computational puzzles, coordinating through group and direct messaging — all running in containers on off-the-shelf harnesses.
The visualization above replays telemetry from messages and shared-workspace file access as the agents solved the game's challenges. The same approach — define responsibilities, exchange findings, inspect what happened, and refine the instructions — can be applied to software engineering, research, and operations.
## Scion architecture companions
Scion is designed to b