content-forge-vault/00-inbox/Subagents.md
2026-03-03 02:00:01 +08:00

1.6 KiB

Subagents

Subagents are Claude Code's way of running parallel workers. Each one gets its own fresh context window — like assigning a task to a new employee who isn't overloaded with everything else you've been working on.

The key insight: subagents solve the Context Window Management problem. Instead of cramming everything into one conversation, you delegate isolated tasks to fresh agents.

The Three Types

  1. Explore agent — reads files, searches codebase, answers questions about code. Read-only.
  2. Code agent — can read AND write files, run commands. Full power.
  3. Custom agent — via the Task tool with specific instructions.

When to Use Them

Boris Cherny runs 5 local sessions + 5-10 web sessions in parallel. That's extreme for most people. The practical version:

  • Research tasks — send a subagent to explore a codebase or read docs while you keep working
  • Independent tasks — two features that don't touch the same files? Parallelize.
  • Context isolation — prevent a debugging rabbit hole from polluting your main session
  • Verbose operations — code reviews, file analysis, anything that generates a lot of output you don't need to see

The Pattern

Each subagent gets:

  • A clear task description
  • Relevant file paths or search scope
  • Expected output format
  • Its own fresh context window

The main session stays clean. You get a summary back.