Skip to content

Glossary

This glossary defines key terms used throughout the Scion documentation and ecosystem.

An isolated worker instance running an LLM harness. Each agent has its own identity, workspace, and configuration.

A project-level grouping of agents and configuration, typically corresponding to a git repository and a .scion directory.

An adapter that allows an underlying LLM tool (like Gemini CLI or Claude Code) to run within the Scion orchestration layer.

The centralized control plane in a hosted Scion deployment. It manages identity, grove registration, and dispatches tasks to Runtime Brokers.

A set of configuration overrides that define how a runtime should execute an agent (e.g., resource limits, environment variables).

The underlying technology used to execute agent containers (e.g., Docker, Podman, Apple Virtualization, Kubernetes).

A compute node that executes agents. It connects to a Hub to receive instructions and reports agent status.

A helper utility bundled with Scion that is injected into agent containers to provide status reporting, metadata access, and task management.

A versioned blueprint for an agent, defining its base image, system prompt, tools, and initial state.

A unique identifier for a grove. Git-backed groves use deterministic UUID v5 identifiers derived from the normalized git URL. Hub-native groves use random UUID v4 identifiers.

An extension module built on hashicorp/go-plugin that provides additional capabilities (e.g., message broker or agent harness implementations) without modifying the Scion core.

A persistent, mutable storage volume shared between agents within a single grove. Backed by host filesystem directories (local) or Kubernetes PersistentVolumeClaims (K8s).

The working directory mounted into an agent container, typically managed as a Git worktree (local mode) or provisioned via git init + git fetch (Hub mode) to ensure isolation from other agents.