🌱

SCION.

Multi-Agent Orchestration Explainer

Scion Core

A five-part containerized orchestration system for multi-agent workflows.

01

Define

Agent Templates

02

Boot

Sandbox Runtimes

03

Spawn

Swarm Hierarchies

04

Notify

Reactive Wakeups

05

Sync

Shared Filesystem

Agent Definition

In Scion, an agent is defined by a structured template consisting of four core layers.

1. System Instruction

A system instruction defining the agent's primary purpose, persona, and core role (e.g. "You are a Security Auditor. Inspect lockfiles..."). Configured as the foundational LLM boundary.

2. Agent Instructions

Detailing how the agent primarily gets its work done. Scion dynamically appends contextual instructions like agents-git.md or agents-hub.md based on its workspace.

3. Agent Skills

A set of specialized skills conforming to the **Agent Skills Specification** (agentskills.io) to perform deterministic CLI tasks (e.g. compilers, test tools, linters).

4. Model Context Protocol (MCP)

Optionally configures MCP server connections to dynamically expose external tools, with underlying container harnesses (Gemini, Claude, Codex) serving as minor, optional execution details.

Agent Runtime

Scion abstracts complex system execution, managing secure, isolated processes across varied backends under a managed control plane.

  • Agent Lifecycle: Uniform control commands (create, start, stop, attach, delete) regardless of the underlying agent implementation.
  • Container Runtimes: Seamlessly executes sandboxed processes across Docker, Podman, macOS native Apple Virtualization, or remote Kubernetes.
  • Env & Volume Mounts: Projects explicitly defined environment variables, mounts read-only secrets/credentials, and maps isolated tmpfs shadow volumes.
  • Native & Federated Identity: Binds native project-scoped identities (project--agent) to federated systems like GCP service accounts or Agent Platform credentials.
Agent State Model Simulator
Dimension 1: Infrastructure Phase
provisioning running stopping stopped
Dimension 2: Cognitive Activity
offline working thinking waiting_for_input blocked
Dimension 3: Detail Stream
Phase: provisioning
Activity: offline
Status log: Allocating container sandbox...
Click to trigger transitions

Agent Collaborators

Scion supports teams of humans working with teams of agents. Both are treated as actors in the system, enabling dynamic, self-directed collaboration instead of rigid, hard-coded orchestration graphs.

  • Actor Interaction: Both humans and agents can query, message, or inspect Agents.
  • Self-Directed Management: Agents are equipped with access to the scion CLI, letting agents themselves dynamically spawn, query, or stop sub-agents.
  • Asynchronous Messaging: Collaborators communicate asynchronously, sending messages and delegating work across safe, pluggable brokers.
  • Interactive Inspection: Humans and parent agents can attach or "look" at running agent processes (via tmux) or view logs.
🧑‍💻 👤 🤖 🤖 🤖 🤖 🤖 🤖 🤖

Agent Notifications

Scion combines state and messages into notifications to provide robust asynchronous event driven orchestration patterns.

  • State & Message Intersection: Notifications represent events triggered when an agent transitions state (e.g. from working to waiting_for_input).
  • Sub-Agent Suspend Pattern: When an agent delegates a task, it invokes a child, updates its activity to blocked, and sleeps. No active polling occurs.
  • Event-Driven Wakeup: Once the sub-agent completes, stalls, or needs input, a notification hook automatically wakes the parent agent up.
  • Human Attention Alerts: Triggers notifications to slack, terminals, or dashboards the instant an agent gets stuck or requires human confirmation.
Reactive Notification Wakeup
🤖 COORDINATOR 🤖 🤖 🤖 🤖 🤖 🤖

Shared Filesystem

For multi-agent workflows, sharing a high-performance filesystem provides frictionless code modification, test execution, and file exchanges. Scion manages concurrent file access safely.

  • Shared File Volume Mounts:
    • Agent workspaces share the same physical project directory, avoiding heavy git clone replication delays and preserving disk space.
  • Pass Context by Reference:
    • Instead of bloating LLM payloads with massive files or logs, agents pass a file path reference on the shared volume, drastically saving token bandwidth and context window size.
  • Shared Learnings & Progress Notes:
    • Sibling agents read and write shared folders to exchange progress summaries, structured learnings, or execution logs in real time, eliminating chat synchronization overhead.
  • Workspace Branch Isolation:
    • If Git isolation is needed, agents can initialize local branches or isolated worktrees, but they still operate on the shared disk storage.
Shared Filesystem Simulator
🤖 Agent 1 🤖 Agent 2 📁 scion/ 📁 src/ 📁 tests/ 📄 auth.py 🔒 📄 server.py 🔒 📄 test_auth.py 🔒
Click "Run Simulation" to observe concurrent filesystem operations.

Scion Orchestration Review

From declarative template definitions to secure, low-latency filesystem execution, Scion coordinates complex agentic workflows in five key stages.

1. Define

Parses declarative agent templates containing system instructions, prompts, custom skills, and MCP configurations.

2. Boot

Launches standard execution sandboxes, isolating credentials, variables, and runtime boundaries dynamically.

3. Spawn

Coordinates human actors and P2P agent hierarchies with lateral peer-to-peer messaging brokers.

4. Notify

Enforces reactive wakeup sleep cycles (suspends parent, wakes instantly on child completion notification).

A1 Client

5. Sync

Shares folder filesystems concurrently to pass large context by reference and exchange shared progress notes.