Skip to content

Sub-agents

A sub-agent is a helper that another agent (its parent) starts for part of a task. It can be Claude Code, Codex or Pi.

The status bar shows Sub-agents with a count. Open it to see the tree and each sub-agent’s state.

  • Choose a row to open its conversation. The page header shows its state, how long it has run and, while it works, the step it is on.
  • When it has finished, its Result shows above the conversation.
  • Type in Prompt this sub-agent… to send an instruction. An idle one resumes. Prompt on a row opens it with the cursor there.
  • Stop ends it.

Sub-agent approvals are grouped under the parent, and each card names the sub-agent.

In Computer settings, Sub-agents, you can:

  • Choose which agents may start sub-agents, and which models and effort levels they may use.
  • Limit how many run at once, how deep they nest and how many tokens each may use.
  • Turn on Terminal agents to apply the same rules to agents you start in a terminal.

A refused start says why, such as “max depth 2 reached”. Panic stop also cancels every sub-agent.