Threads
Delegating to child threads
Let an agent start child threads on other harnesses and wait for their results.
An agent can split work across child threads, each on any harness and model you have connected, and wait for their results. A Claude Code thread can ask Codex for a review, or several children can try different approaches at once. Each child has its own workspace and native history and reports its result to the parent.
How it works
Ask for it in plain words ("have Codex review this in a child thread") and the agent uses its delegation tools:
| Tool | What it does |
|---|---|
spawn_task | Start a child thread with a prompt, harness and model. |
wait_for_tasks | Wait for up to 8 children, for at most 240 seconds per call. |
read_task | Read a child's transcript. |
get_task_result | Read a child's saved result. |
send_task_message | Send a child a follow-up. |
stop_task, resume_task | Stop or resume a child. |
find_tasks, list_task_agents | Find other threads and the agents available. |
A family of threads runs at most 8 child launches at once and nests at most 3 levels deep. Files move between threads through saved results and the project source, never through another agent's native conversation.
Lineage
Thread details → Lineage shows a thread's parent, children and forks; the thread menu links to the parent. Each child appears in the parent's transcript as a card with its harness, model, status and Open thread. Child threads use credits like any thread.
Bring another thread in yourself
@in the composer has a Threads group: attaching a thread adds its saved history as context.- Fork thread starts a new thread from any finished turn. The original stays unchanged; harnesses that support native forks keep native history, others receive the saved context.
- Switching a thread to another harness is a lossy handoff, marked in the transcript.