Anthropic's official guide: how and when to use subagents in Claude Code
daniel_mac8 · x · 2026-10-25
Anthropic published an official best-practices article, "How and when to use subagents in Claude Code" (17-min read). Core idea: long sessions accumulate weight — every file read, tangent, and half-finished thought stays in the context window, slowing responses and raising token costs. A subagent is an isolated Claude instance with its own context window: it takes a task, works independently (reading files, exploring code, making changes), and returns only the result — like a browser tab for your Claude Code session, letting you chase tangents without losing the main thread. Multiple subagents can run in parallel, each with different permissions (e.g., a research subagent with read-only access). The article covers when delegation is worth it, how to invoke subagents, and when the overhead isn't.
Related event: Anthropic releases guide on when and how to use Claude Code subagents(2 posts)→
More from coding & agent
- After a week under active cyberattack, an engineer explains why software factories are now essential — danielrock · 2026-10-25
- Bots talking to bots: one agent auto-contacts Guo Yu, whose bot answers on his behalf — huangyun_122 · 2026-10-25
- Free open-source AI engineering course builds every algorithm from raw math before frameworks — ghumare64 · 2026-10-25
- "No human will ever write code better than LLMs," declares engineer Arpit Bhayani — arpit_bhayani · 2026-10-25
- Dev launches staygreen.io, a lifetime-free uptime monitor, after balking at $330/mo white-label pricing — PawelHuryn · 2026-10-25
- ComfyUI plugin turns Load Image into a mini Photoshop with editable OpenPose — Occsan · 2026-10-25