Chapter 10: Sub-Agent Orchestration — let multiple AI avatars work in parallel
Sub-agents let you set up several additional models that assist the main model as sub-agents.
Reading time: about 10 minutes.
10.1 What are sub-agents?
When a task contains at least two independent sub-tasks, the AI can dispatch "clones" (sub-agents) to work in parallel, e.g.:
- Research two technical approaches at the same time;
- Review two pieces of code at the same time;
- Design two candidate solutions and compare them.
Principles:
- Simple answers, reading known files, single edits, and dependent sequential steps are NOT suitable for sub-agents;
- Sub-agent results need review — never treat them as final conclusions directly;
- Sub-agents have limited powers: no commands, no shared-storage changes, no out-of-scope writes.
Beginners don't need to manage this — when a task truly suits parallel work, the AI decides whether to spawn clones on its own.
10.2 Configuring sub-agents
Configure sub-agents in the Settings page. For each sub-agent you can choose:
- The model it uses
- Its provider
- A system prompt (defining the sub-agent's role and responsibilities)
10.3 Using them in chat
Once configured, open the bottom action panel on the chat page to enable or disable sub-agents.
10.4 How they work
- When you give a complex copy task, the main model can distribute sub-tasks to sub-agents.
- Multiple sub-agents can process different sub-tasks in parallel, and the main model then aggregates the results, helping finish the overall task faster.
Note
Giving different sub-agents different roles and models lets each do its own job — e.g. one sub-agent handles code review while another handles documentation.
End of chapter · Next: Agent Tools — the "hands" the AI can use