Subagent conversations
Found this helpful? Share it:
Found this helpful? Share it:
Sometimes one agent isn't enough. With subagent conversations, an agent can hand off parts of a task to focused worker chats while it keeps coordinating the main response 🤝
Subagent lets an agent create a worker conversation. The worker uses the same agent identity as the parent chat, but it gets its own thread for the delegated work.
For example, the parent can ask one worker to research sources while another drafts a summary. The parent then combines the results in the main chat.
Subagent is a per-chat toggle. Turn it on in the composer tools dropdown:
Open the tools dropdown (the sliders icon next to the message box)
Find the Subagent row and turn it on
When enabled, the AI can create worker conversations from the current chat as needed. See Chat settings and toggles for the rest of the controls in that dropdown.
Worker chats stay linked to the parent chat, so you can review them from the subagent card in the conversation.
When the parent agent decides it needs help, it creates a child conversation and sends the worker a clear request. The worker runs independently and reports back with results.
Worker chats do not appear in your normal recent chats list. Open them from the parent chat's subagent card. When you are inside a worker chat, use the parent shortcut at the top of the chat to get back.
The parent agent waits for the child to finish, then incorporates the results into its own response.
When the parent hands off a task, it can shape how each worker runs:
Run several at once: the parent can fan out workers in the background and keep going, then gather their results as they come back, instead of waiting on one at a time.
Pick a model per worker: a worker can run on a different model than the parent, so heavy research can use a capable model while quick steps use a fast one.
Sandbox a worker: the parent can put a worker in a read-only mode so it can look things up and report back without changing anything.
If a worker needs to run an action that requires your approval, the confirmation surfaces in your chat, so you stay in control even though the work is happening in a linked conversation. See Autonomy and confirmations.
Worker chats inherit bounded execution limits from the parent run: recursion depth, child-call count, per-response output tokens, and total output tokens. The parent can narrow those limits when it delegates work.
Child work still counts toward the same account-level usage and credit controls, so delegation cannot bypass payment or rate limits.
To keep workflows predictable and costs controlled, there are a few guardrails:
No infinite chains, child agents can delegate only within the enforced recursion-depth and child-call limits.
Same agent identity, workers run as the same agent as the parent chat.
Linked access, worker chats are reached from the parent chat, not from the main recent chats list.
These constraints keep your workflows simple and your costs visible.
Related articles
Was this helpful?