{
  "id": "definition:subagent.canonical",
  "_kind": "Definition",
  "_file": "terminology/definitions/subagent-canonical.yaml",
  "_cluster": "terminology",
  "attributes": {
    "authoredAt": "2026-04-28T00:00:00Z",
    "termId": "term:subagent",
    "text": "A specialized worker / persona that an agent runtime can dispatch alongside\nits primary loop. Canonical term follows Anthropic Claude Code (\"subagent\",\none word; docs.anthropic.com/en/docs/claude-code/sub-agents). The same\nNodeKind covers two provenance variants — vendor-shipped built-ins (Claude\nCode's general-purpose / Explore / Plan; OpenAI built-in handoff patterns)\nand user-installable custom definitions (Claude Code `.claude/agents/<name>.md`,\nparallel to a Skill's SKILL.md). Distinguished via the `provenance`\nattribute. The dispatch *mechanism* a runtime uses to invoke a subagent is\nrecorded on `AgentRuntimeImpl.subagentDispatchMechanism` (task-tool /\nhandoff / sub_agents-array / agent-as-tool / none); the *definition format*\na platform loads custom subagents from is recorded on\n`AgentPlatformImpl.subagentDefinitionFormat`.\n",
    "status": "canonical"
  },
  "outgoingEdges": [
    {
      "from": "definition:subagent.canonical",
      "to": "term:subagent",
      "kind": "defines"
    }
  ],
  "incomingEdges": []
}