Developing Subagents

Subagents isolate verbose work from parent context, encode repeated workflows, and route narrower tasks to cheaper models. Prefer orchestration from the main conversation; nested delegation is platform-specific, harder to predict, and more expensive. Compose reusable workflow knowledge with Skills.

Platforms

• Claude Code — .claude/agents/, frontmatter fields, permission modes, hooks, skills, memory, MCP scoping
• OpenCode — .opencode/agents/, primary/subagent modes, per-tool allow/ask/deny, cascade rules, JSON form
• Codex — .codex/agents/, TOML custom agents, explicit parallel spawning, inherited sandbox/config, global [agents] limits

When to make a subagent

• A workflow keeps repeating with the same instructions — encode it once.
• A task floods main context (large searches, log processing, full-file reads).
• The job needs a different system prompt, model, or permissions than the parent.

If a Skill carries the same instructions, prefer the Skill — Skills compose across agents and platforms.

Description: selection metadata