Google Open-Sources Scion Multi-Agent Orchestration Testbed

Google open-sourced Scion on March 27, 2026, introducing an experimental multi-agent orchestration testbed designed to manage concurrent LLM-based agents. The framework enables developers to deploy groups of specialized agents that operate within isolated containers across local machines, remote virtual machines, or Kubernetes clusters.

Scion focuses on the management of deep agents, such as Gemini CLI, Claude Code, and OpenAI Codex. By providing each agent with its own isolated identity, credentials, and workspace, the system allows multiple agents to collaborate on a single project simultaneously without interfering with one another’s processes.

This isolation is achieved by assigning each agent its own container and git worktree. This architecture supports the parallel execution of diverse technical tasks, including software coding, security auditing, data research, and testing, where different agents can tackle separate components of a codebase in tandem.

Manager-Worker Architecture and the Grove

Scion utilizes a Manager-Worker architecture to separate the orchestration layer from the agent execution environment. The host-side CLI serves as the manager, orchestrating the lifecycle of the agents and managing the Grove, which is the designated project workspace.

The workers are the isolated runtime containers, such as Docker, that execute the agent software. This separation makes Scion harness-agnostic, meaning the orchestration layer remains independent of the specific model or agent tool being used. Developers can mix and swap different agents—such as combining a QA agent with a coder and a security auditor—without altering the underlying orchestration infrastructure.

To manage these environments, Scion employs a configuration system based on Profiles, Runtimes, and Harnesses. This allows users to switch easily between different environments, such as moving from local Docker containers to remote Kubernetes clusters. Configuration is handled through two levels of settings: user-wide defaults located at ~/.scion/settings.yaml and project-specific overrides found in .scion/settings.yaml.

Operational Workflow and Tooling

The framework provides a set of CLI tools for developers to initialize and interact with their agent swarms. The process begins with the scion init command, which creates the necessary .scion directory within the project root.

Once initialized, developers can launch specific agents using the scion start <agent-name> “<task>” command. The system also supports human-in-the-loop control through tmux sessions, allowing developers to monitor and intervene in agent activities. Other operational commands include scion attach for interacting with an active session, scion logs for viewing output, and scion resume for restarting stopped agents while preserving their state.

For monitoring and analysis, Scion uses OpenTelemetry (OTEL) to provide normalized telemetry across different harnesses. This allows developers to track messages and file access within the shared workspace.

Orchestration Philosophy

Scion departs from traditional multi-agent frameworks that rely on rigid protocols or complex, predefined orchestration structures. Instead, it adopts a less is more approach, acting as a hypervisor for agents rather than a full-stack framework.

Under this philosophy, the orchestration layer provides the foundation—memory, containers, and workspace isolation—but leaves the coordination to the models themselves. Agents dynamically learn to use a CLI tool to communicate and coordinate their tasks through natural language prompting. This approach is based on the premise that as frontier models become more capable, the need for explicit, hard-coded orchestration structures decreases.

To demonstrate this capability, Google highlighted Relics of Athenaeum, an agent game where multi-agent orchestration is defined entirely in markdown. In this scenario, a group of agents collaborates to solve computational puzzles by coordinating through direct and group messaging, all while running in containers on standard harnesses.

As an open-source project, Scion currently requires users to build images from source, which necessitates the installation of golang. The project is positioned as a rapid prototype testbed for experimenting with evolving multi-agent patterns in software engineering and platform engineering workflows.