Running one AI coding agent is straightforward. Running five of them at once, each with its own credentials and workspace, without them stepping on each other? That's the problem Google Cloud Platform built Scion to solve.
Scion is an open-source testbed for orchestrating multiple LLM-based agents concurrently. Each agent runs in its own container (Docker locally, Kubernetes for remote clusters) with isolated credentials, workspaces, and identities. You could spin up a Gemini CLI agent handling research, a Claude Code agent writing backend code, and an OpenAI Codex agent running tests, all working on the same project simultaneously without conflicts.
The architecture follows a manager-worker pattern. The CLI acts as a host-side manager controlling agent lifecycles and a shared project workspace called the "Grove." Workers are the containerized agents themselves. Configuration is handled through three concepts: Profiles (environment settings), Runtimes (Docker vs. Kubernetes), and Harnesses (which agent software to run). Agents can be paused and resumed with their full state preserved, and you can attach to any running agent to monitor its logs or interact directly.
This is clearly an experimental project, not a polished product. Google labels it a "testbed," and the documentation reflects that. But the underlying idea is practical. As AI coding assistants get more capable, the bottleneck shifts from "can the agent do the task" to "how do I coordinate several agents working on different parts of a codebase without them overwriting each other's files or using the wrong API keys."
For developers already juggling tools like Claude Code, Cursor, or Gemini CLI, Scion offers a structured way to run them side by side. It's early-stage and requires comfort with Docker and CLI tooling, but the isolation model is sound. If multi-agent workflows are part of your development setup, or you're curious about where that trend is heading, Scion is worth a look on GitHub.