Sub-Agent

A sub-agent is an AI agent created by another AI agent — the "parent" — to handle a specific piece of a larger task. The parent agent delegates work to sub-agents, collects their results, and coordinates the overall effort.

Why sub-agents exist

Every AI model has a context window — the maximum amount of text it can process at once. When a parent agent is handling a complex task, its context window fills up with instructions, conversation history, and intermediate results. Instead of cramming everything into one overloaded agent, you spawn sub-agents that each get their own fresh context window and a focused brief.

This is the AI equivalent of a manager delegating work. A marketing director does not write every blog post, design every graphic, and send every email personally. They delegate to specialists. Sub-agents work the same way.

How sub-agents work

1. The parent agent receives a complex goal
2. It breaks the goal into sub-tasks
3. For each sub-task, it spawns a sub-agent with a specific brief
4. Each sub-agent executes its task independently
5. Results flow back to the parent agent
6. The parent synthesises and delivers the final output

Spawning patterns

• Fan-out / Fan-in: Parent spawns multiple sub-agents in parallel (e.g., researching 5 competitors simultaneously), then collects all results
• Pipeline: Each sub-agent passes output to the next (research → write → edit → format)
• Specialist pool: Parent routes different task types to the appropriate specialist sub-agent

In practice

In Claude Code, the Agent tool spawns sub-agents. Each sub-agent has access to the same tools as the parent but operates with its own context. You can set `isolation: "worktree"` to give a sub-agent its own copy of the codebase for safe parallel development.

Key consideration: Every sub-agent uses tokens. A task that spawns 5 sub-agents costs roughly 3-5x more than a single agent doing everything sequentially. Use sub-agents when the quality or speed benefit justifies the additional cost.