Claude Subagents

Claude Subagents are specialized assistants inside one Claude Code session. Official docs say to use one when a side task would flood the main transcript with search hits, logs, or file dumps you will not need again. The subagent works in its own context window with a custom system prompt, tool access, and permissions, then returns a summary. Define a custom subagent when you keep spawning the same kind of worker.

Docs distinguish subagents from larger parallelism: background agents / agent view for many independent sessions, cross-session messaging, and agent teams. Do not treat the old "256K windows and 90%+ gains" line as current first-party spec. Use the live subagent guide.

Key Features

• Isolated context: Research stays out of the parent thread.
• Custom prompt, tools, permissions: Restrict what a worker can touch, including MCP scope.
• Built-in types: Docs still describe built-in subagents plus user-defined ones.
• Persistent memory option: Frontmatter can enable memory for repeat workers.
• Single-session scope: For many sessions on one screen, use agent view, not more subagents.

Use Cases

• Codebase explore that would otherwise dump hundreds of files into chat.
• Constrained implementers that may edit only a directory.
• Review workers that return findings without keeping the raw diff in the parent.

Pricing

Subagents consume the same Claude Code session and model usage. Extra workers cost extra tokens. There is no separate subagent SKU. Recheck claude.com/pricing.

Getting Started

1. Read Create custom subagents.
2. Add a subagent file in the scope the docs describe (project or user).
3. Give it a tight tool list before you let it edit.
4. If you need many parallel checkouts, switch to worktrees or agent view.

Frequently Asked Questions

Is this a multi-agent product?

It is workers inside one session. Teams and background agents are separate official features.

Can I pin a model per subagent?

Yes. The configure section includes choosing a model.

90% faster?

That was earlier marketing in our old entry. Official docs now talk about context savings, not a single percentage.

Alternatives

• Claude Skills: Reusable procedures, optionally run inside a subagent.
• Claude Code: Agent view and teams for bigger parallelism.
• Codex CLI: OpenAI's own subagent/session model.

Tips

1. Return summaries only. If the parent needs the file, the worker failed.
2. Preload a skill into a subagent instead of pasting a novel into its prompt.
3. Recheck built-in names. Anthropic has been renaming explore/plan workers.

Conclusion

Subagents keep Claude Code's main thread readable. Start from the subagent docs, isolate one research job, then add permissions.