Sub-agents

A sub-agent is a helper agent that another agent starts to do part of a task. The agent that starts it is its parent. A sub-agent can be Claude Code, Codex or Pi.

Follow sub-agents

When a session has sub-agents, the status bar shows Sub-agents with a count. Open it to see the tree. Each row shows its state: Starting, Running, Waiting for approval, Idle or Ended.

• Choose a row to open that sub-agent’s conversation.
• Type in Prompt this sub-agent… to send it an instruction. An idle one resumes.
• Choose Stop to end it.

Approvals from sub-agents reach you grouped under their parent, for example “3 sub-agents waiting”. Each card names the sub-agent.

Set the rules

Open Computer settings and choose Sub-agents.

• Pick which agents may start sub-agents, and which models and effort levels they may use.
• Set limits: how many at once per agent and on the computer, how deep they can nest, and a token budget per sub-agent and per agent.
• Terminal agents applies the same rules to agents you start in a terminal.

A refused start tells you why, for example “max depth 2 reached” or “model not in the Codex allowlist”. Panic stop also cancels every sub-agent.