# Scion
Scion is an open-source multi-agent orchestration testbed from Google Cloud Platform that runs multiple AI coding agents in parallel — each in its own container, with its own git worktree and credentials — collaborating on the same codebase without stepping on each other.
It is harness-agnostic: [[Claude Code]], [[Gemini CLI]], [[OpenCode]], and [[OpenAI Codex|Codex]] all plug in. Distribution is Apache 2.0. Explicitly *not* an officially supported Google product — it's an experiment Google publishes to push the field.
## The problem
Running a single coding agent on a repo is straightforward. Running *several* — different harnesses, different roles, different files in flight — falls apart fast: shared filesystem state, conflicting branches, credential collisions, observability blind spots. Most existing "multi-agent" demos are scripted choreography; Scion bets that real multi-agent work needs real isolation primitives.
The wager: give each agent a container + a [[Git Worktree]] + scoped credentials, then let the models themselves figure out how to coordinate via natural language. "Less is more" — no rigid orchestration framework imposed.
## How it works
- **Containerized agents.** Every agent runs in its own container — [[Docker]], [[Podman]], Apple Container Runtime, or [[Kubernetes]] — with its own credentials and environment.
- **Git worktrees per agent.** Each agent works in its own worktree off the shared repo, so concurrent edits don't collide at the filesystem level. Merging is a deliberate step, not an accident.
- **Tmux-based sessions.** Attach/detach to live agent sessions; agents keep working in the background. Remote tunneling is supported.
- **Templates for specialization.** Define agent roles via system prompts + skill sets — debugger, reviewer, implementer, etc.
- **Optional Hub.** A central control plane for managing swarms across machines; Runtime Brokers expose container runtimes to the Hub for distributed execution.
- **Normalized OTEL telemetry.** Observability across the swarm in a unified format — non-trivial when agents come from different vendors with different log shapes.
## Usage shape
```bash
go install github.com/GoogleCloudPlatform/scion/cmd/scion@latest
scion init --machine # initialize the machine
cd my-project && scion init # initialize a project
scion start debug "Help me debug this error" --attach
scion list # list active agents
scion message <name> "..." # send a message to an agent
scion logs <name> # tail logs
```
## Design choices worth noting
- **Isolation as a first-class primitive.** Container + worktree + scoped creds is the right shape for "many agents, one repo." Anything less re-introduces the conflicts the tool exists to prevent.
- **Coordination by prompt, not by orchestrator.** Most multi-agent frameworks bake in an opinion about how agents should hand off, vote, or critique. Scion punts that to the models. Risky on weak models, freeing on strong ones.
- **Harness-agnostic, deliberately.** A bet that no single agent harness wins, and that real workflows mix at least two.
- **Tmux as the human interface.** Attach when you want to look, detach when you don't. Familiar shape for terminal-native users.
- **Local-first with optional Hub.** Most use is single-machine; Hub is there when you outgrow that, not before.
- **Honest status disclosure.** Local mode "relatively stable," Hub workflows "~80% verified," Kubernetes runtime "early." This is alpha software shipped with adult labelling.
## Where it fits
For anyone exploring how to actually use multiple coding agents on the same project — the legitimate version of "AI swarms," not the demo version — Scion is one of the more credible primitives. Adjacent to [[OpenClaw]] (personal-AI orchestration spanning channels and devices) and [[AskOracle]] (multi-model fan-out for *prompts*, not parallel-execution agents). Different layer of the stack: Scion is "many agents working in a repo right now"; AskOracle is "many models answering one question once."
## References
- Source: <https://github.com/GoogleCloudPlatform/scion>
- License: Apache 2.0
- Maintainer: Google Cloud Platform (community / experimental — not an officially supported Google product)
## Related
- [[AI Agents]]
- [[Claude Code]]
- [[Gemini CLI]]
- [[OpenAI Codex]]
- [[OpenCode]]
- [[Git Worktree]]
- [[Docker]]
- [[Podman]]
- [[Kubernetes]]
- [[TMux]]
- [[OpenClaw]]
- [[AskOracle]]
- [[Superset]] — desktop IDE equivalent of the same wager (worktree-isolated parallel agents, single-machine)