Scion is an experimental multi-agent orchestration testbed designed for running multiple AI agents concurrently within isolated containers. It targets engineers and researchers seeking to prototype and experiment with complex AI-driven workflows, enabling collaborative code development and project management by allowing agents to dynamically learn and coordinate via natural language prompts.

How It Works

Scion orchestrates "deep agents" (e.g., Claude Code, Gemini CLI) as independent, containerized processes. Each agent operates within its own isolated environment, complete with a dedicated workspace, git worktree, and credentials, preventing conflicts. Coordination is achieved through a "less is more" approach where agents dynamically learn CLI tools, enabling them to decide how to collaborate via natural language prompting, fostering rapid prototyping of multi-agent patterns.

Quick Start & Requirements

Installation requires Go: go install github.com/GoogleCloudPlatform/scion/cmd/scion@latest. Users must build their own container images as pre-built binaries are unavailable. Setup involves initializing the machine (scion init --machine) and then the project grove (scion init). Default runtimes (Docker, Container) are auto-detected but configurable. Recommended: add .scion/agents to .gitignore. Official documentation is available for comprehensive guides.

Highlighted Details

• Harness Agnostic: Adaptable to agents running in containers, including Gemini CLI, Claude Code, and Codex.
• True Isolation: Each agent runs in a separate container with distinct credentials and a dedicated git worktree.
• Parallel Execution: Multiple agents operate concurrently as independent processes, locally or remotely.
• Attach/Detach: Agents run in tmux sessions, allowing background operation and interactive attachment.
• Specialization: Agent roles can be defined using Templates with custom system prompts and skills.
• Multi-Runtime: Supports Docker, Podman, Apple containers, and Kubernetes.
• Observability: Integrated OTEL telemetry for logging and metrics across agent swarms.

Maintenance & Community

This project is explicitly labeled "early and experimental," with core concepts settled but "rough edges" expected, particularly in the Kubernetes runtime. It is not an officially supported Google product and is ineligible for the Google Open Source Software Vulnerability Rewards Program. No community channels (e.g., Discord, Slack) are listed in the README.

Licensing & Compatibility

The project is licensed under the Apache License, Version 2.0, which generally permits commercial use and modification.

Limitations & Caveats

The project's experimental nature means features may be incomplete, subject to breaking changes, and future direction is not fixed. Kubernetes runtime support is noted as having known rough edges. Users must build their own container images, and the project lacks official Google support.