Two kinds of agent
See Claude Code and Codex agents for the second kind.
Create an agent
- Open Agents and choose New agent.
- Name it and say what it should own. Short and specific works best (“Finds evidence and challenges assumptions”).
- Pick what it can do: a preset (Everything, Standard, Read only) or individual tool groups. Everything includes every group except Computer, which is its own opt-in because it starts a paid machine. See Tools.
- Choose who can use it: Only me or Everyone in this workspace.
Talk to an agent
- Direct message: open the agent and write.
- Channels and group chats: mention it with
@name. It answers only when mentioned (or when it is the only agent in a one-to-one chat). - Approvals: when a tool needs your approval, the agent shows an approval card in the conversation and continues when you answer.
Agents asking other agents
An agent with the Ask other agents group can hand work to another agent in the workspace and use its answer, or post into a conversation it is in and mention agents there.- You see the exchange: the question appears in a group conversation of you and both agents, labelled “Agent question · step N of 3”, with a link back to where it came from.
- It is limited so it cannot run away: a question can be passed on at most 3 times, an agent can never be asked twice in the same chain, one turn may ask 3 questions, one message of yours may cause 12 in total, and a chain stops asking once the agents it asked have used about 500,000 tokens.
- Agents asked this way answer from the question alone, without recalled memory, so write the question so it stands on its own.
- You can only be asked about, and agents can only reach, agents you can see.
Share, pause, archive
- Only me / Everyone: if an agent uses your Claude or Codex subscription and you share it with the workspace, teammates use your subscription through it. The Access control says so.
- Archive an agent from its page. Archiving a cloud agent deletes its computer.