Google has unveiled Scion, an open-source project serving as an experimental orchestration testbed for managing concurrent agents in distributed computing environments. The system enables developers to run multiple specialized agents—each with its own unique identity, credentials, and workspace—across local machines, remote virtual machines, or Kubernetes clusters.
A 'Hypervisor' for Agent Systems
Google describes Scion as a “hypervisor for agents.” The platform integrates memory management, communication protocols, and task scheduling as orthogonal concerns. This design allows developers to run deep agents like Claude Code, Gemini CLI, or Codex as independent, concurrent processes.
The core strength of Scion lies in its isolation. Every agent operates within its own container, Git worktree, and credential set, ensuring they do not interfere with one another while working on different parts of a project. This architecture supports dynamically evolving task graphs, enabling agents to execute disparate goals—such as coding, auditing, and testing—in parallel. Depending on their lifecycle requirements, Scion supports both long-running, dedicated agents and ephemeral agents designed for single-task execution.
Regarding security, Scion adopts an “isolation-first” strategy. Google notes that the platform does not rely on constraining agent behavior through context-embedded rules. Instead, it grants agents high degrees of freedom to complete tasks while enforcing boundaries at the infrastructure level through container isolation, network policies, and Git worktree restrictions. Scion even encourages a so-called “--yolo” mode, which maximizes execution efficiency while maintaining strict environmental isolation.
To ensure compatibility with the existing ecosystem, Scion utilizes “harnesses”—adapters that manage agent lifecycles, authentication, and configuration. The platform currently supports Gemini and Claude Code, with partial support for OpenCode and Codex. Developers can select from various runtime environments, including Docker, Podman, or Apple containers, via named configuration files.
To demonstrate Scion’s practical capabilities, Google released a codebase for a game titledRelics of the Athenaeum. In this project, multiple agents collaborate via different harnesses to solve computational puzzles. A game runner handles the spawning of new agent roles, which in turn dynamically generate sub-agents to execute specific tasks. These agents synchronize information and collaborate through shared workspaces, direct messaging, and broadcast mechanisms.
