Sub-agents
The agent tool launches a sub-agent: a new chat tab that receives a
task, works it to completion with its own tools and context, and returns its
final answer as the tool result of the parent chat.
When the model uses it
Section titled “When the model uses it”The tool description tells the model to delegate self-contained work and to pass all necessary context in the query — the sub-agent shares nothing with the parent conversation. Typical uses: a wide exploration of an unfamiliar area, a mechanical refactor across many files, or a parallel investigation while the parent keeps reasoning.
What it takes
Section titled “What it takes”| Argument | Meaning |
|---|---|
query |
The task, with all required context |
title |
A few words naming the new tab |
files |
Paths to inject into the sub-agent’s context |
tools |
Built-in tools or tool engine names to grant |
model |
Optional model override |
What you see
Section titled “What you see”A new tab opens with the sub-agent’s chat; switch to it (alt+;) to watch or intervene — it is an ordinary chat, so its tool calls are reviewed the same way. When it produces a final answer the parent’s tool call resolves and the parent turn continues.
Launching a sub-agent is a manual-review tool call: you see the query, title and grants before it starts.