Code

Following the agent's sub-agents

The coding agent can hand part of its work to a sub-agent: a helper agent that works on one piece, such as exploring the code or reviewing a change, and reports back to the agent that started it. A session shows its sub-agents in three places:

  • the sub-agents chip, beside the session's other chips at the top of its page;
  • the Sub agents tab of the rail, which lists every sub-agent of the session;
  • cards in the conversation, one each time a sub-agent starts, reports back, fails or stops.

A sub-agent's own step-by-step work is not in the conversation: the conversation shows the cards, and the Sub agents tab shows everything the sub-agent did. For the rest of the session page, see Working in a session.

The sub-agents chip

The chip saysWhat it means
2 subagents runningTwo sub-agents are working now. The number moves as they start and finish. Point at the chip to see their names.
No subagents runningThe agent can start sub-agents, and none is working now. Point at the chip to see how many ran in this session.
No subagents availableThe agent in the session's latest run can't start sub-agents at all. Point at the chip to see why.

The chip isn't shown until a run has started. Choose it to open the Sub agents tab.

Claude and Codex

Claude and Codex start sub-agents differently. Claude decides for itself when a sub-agent helps, at every effort level. Codex starts sub-agents only when it is asked to: in your message, in the repository's AGENTS.md, or in a skill. A Codex session where nobody asked reads No subagents running, which is expected. To have Codex use sub-agents, ask for them in your message, for example "use a sub-agent to review the migration".

Following a sub-agent

To follow a sub-agent

  1. Choose the sub-agents chip, or in the rail, choose Sub agents. Or choose Open on a sub-agent's card in the conversation.
  2. In the list, choose the sub-agent. Running sub-agents come first, then the newest. A sub-agent started by another sub-agent is indented under it.
  3. Read its status (Running, Reported back, Failed or Stopped), when it started, how long it ran and what it used.
  4. Read its report, what went wrong, or why it stopped.
  5. Under Settings, read the model, the effort, the tools and its full instructions, as far as the agent tells them. inherited means the sub-agent used the main agent's setting; observed means the tools it was seen using.
  6. Under Conversation, read what the sub-agent did. It updates while the sub-agent works.
  7. To go back to the list, choose the tab's title.

A sub-agent's conversation is read-only: you can't send a sub-agent a message, stop it or change its settings. To change what it does, message the session's agent.

When a sub-agent stops

A sub-agent shows Stopped when you stop the run, when the run ends while it is still working, or when the agent stops it. Failed means the sub-agent itself hit an error.

The sub-agents of a session stay in its Sub agents tab after they finish and after you reload the page.

Last modified on