AI/Agent/Subagent
Subagent
- A subagent is a specialized agent definition with its own prompt, tools, model selection, context, or execution policy, which the main agent delegates to. The point is context: the side task runs in its own window and returns a summary instead of flooding the main conversation.
- Like AI/Agent/Hook and unlike Agent/Skills, there is no open specification — only each vendor's reference.
Why subagents do not port
- The file format nearly converges and then doesn't. Markdown with YAML frontmatter in an
agents/ directory covers Claude/Code/Subagent, CursorAI/Subagent, GeminiCLI/Subagent, GitHub/CoPilot/CLI/Subagent (as *.agent.md), and OpenCode/Subagent — but Codex/Subagent uses TOML in ~/.codex/agents/, and PiAI/Subagent has no declarative format at all. - The real gap is capability. A Claude definition declares
model, effort, maxTurns, permissionMode, memory, background, hooks, and isolation: worktree; Cursor documents a much smaller frontmatter surface; Gemini's are preview. Fields drop silently on the way across. - Standalone host configuration and packaged plugin composition are separate surfaces. Codex documents subagents fully and its plugin manifest still has no
agents field, so a Codex/Plugin cannot ship one. "Does tool X support subagents" is the wrong question — ask which surface. See AI/Coding/Tool/Plugin/Report/26/Agentic Coding Plugin Systems and Cross-Provider Portability.
- Related but distinct: AI/Agent/Multi is about orchestrating agents; a subagent is the unit being orchestrated.