Skip to main content
For a big job, an agent can split the work and hand parts of it to subagents: fresh, isolated copies of itself that each run on their own and report back. The agent stays in charge, gathers the results, and gives you one answer.

​
Why it helps

A single agent works through a task one step at a time. Subagents let it work on several fronts at once.
  • Parallel work. Three independent lookups run at the same time instead of one after another, so the whole job finishes sooner.
  • A clean workspace. A large or noisy subtask (reading a long thread, crawling a site, sifting a data dump) runs in the subagent’s own context, so it doesn’t crowd out what the main agent is holding in mind.
  • A focused helper. A subagent can take on an extra persona for its task, so one part of the work runs with a specialist’s framing while the rest stays general.
Each subagent is the same agent, with the same memory, connected services, and identity. It just runs on its own for one scoped task.

​
What it looks like

When an agent delegates, a card appears showing the work in flight. This works in both web chat and the skydive chat CLI. Its header reads Subagent working for a single helper, or Subagents working in parallel for a fan-out. Each subagent is one row, named for its task.
In web chat, expand the card to watch the rows or dismiss it. In the CLI, the card streams live in the transcript. Each row shows the task name, a live activity line, and an elapsed time and context-token count as the subagent works. Use ↑ / ↓ to select a row and press ↵ to open that subagent’s conversation directly. esc returns you to the parent chat. Fan-outs wider than eight rows collapse into a ↓ N more tail while running. As each subagent finishes, its result flows back to the main agent. Once the batch settles, the transcript keeps a short Subagents finished running divider where the work happened, and the rows remain navigable so you can review what each helper did.

​
When agents use it

The agent reaches for subagents when the shape of the work fits, not on every task:
  • Several independent pieces that can run at once (research across three sources, checks against multiple systems).
  • A heavy subtask worth keeping out of the main thread (a long document to read, a large result set to distill).
  • A part of the job that benefits from a specialist framing.
For a quick, single-threaded task, the agent just does it inline. No card appears.

​
Cost

Each subagent is a full agent run and counts toward usage like any other run. A well-scoped subtask (a search, a summary, a mechanical edit) often runs on a lighter, cheaper model, so a fan-out of small jobs stays inexpensive. See Usage for how runs are metered.

See work across conversations

The feed shows what your agents are doing everywhere, not just the current chat.