Big jobs go faster and stay cleaner when split across focused helpers working in parallel.
Some requests contain many independent pieces: "research these 12 companies". One agent doing all of it gets slow, and its conversation fills up. Subagents let an agent act as a coordinator that delegates.
Two kinds of helper:
- A self-clone (the "(Me)" entry, on by default): a copy with the same instructions and tools. Ideal for splitting one kind of job into parallel parts, one company each. Clones can't clone themselves again.
- A named specialist: another agent you add under Subagents, with its own instructions, connectors and skills, such as a "Finance analyst" agent.
Each subagent works in its own conversation and its own sandbox, and the coordinator combines their results into its reply. Several can run at the same time (up to 10 per task), and they can work in the background.
Setting up a specialist:
- Build and test the specialist first, with a clear name and description.
- Attach it to the coordinator.
- In the coordinator's instructions, say when to use each specialist and what result to expect back.
- Make sure people who use the coordinator can also access the specialist.
Practical notes: files aren't copied automatically, so the coordinator must pass along what a subagent needs. Subagents act with the permissions and accounts of the person who started the task, and they can pause for approvals too. Each subagent's work uses credits, visible in the task's breakdown.
Skills or subagents? A skill teaches one agent how to do something; a subagent is a separate worker with its own context.
Remember
- A coordinator can delegate to self-clones or named specialists
- Subagents run in parallel, each with its own context and sandbox
- Pass files explicitly and describe in instructions when to use each helper