One assistant holding every tool looks simple until you use it. Each request drags the whole toolbox along, the model burns tokens deciding what to ignore, and answers slow down. Splitting the work across a roster of small specialists fixes both problems. Each one gets a narrow job, a short tool list and a matching reasoning level.
What it is
A sub-agent is a separate worker with its own instructions, its own permissions and its own model setting. A main agent reads the request, picks the right specialist and hands off. The specialist does one thing and returns a result. Because each worker carries less context, a request that once cost a large bill can run for a fraction of it.
The roster matters more than any single agent. A typical set covers research, fact checking, writing, data cleanup, scheduling and review. You do not need fifteen on day one. Start with the three tasks you repeat most.
When to use it
Reach for this when your current assistant is slow, pricey or unreliable on a specific job. It also helps when one task needs strict limits, such as an agent that may read files but never send messages. If you run a single simple prompt a few times a day, stay with one agent.
How to set it up
- List your repeated tasks and group them by the tools each one truly needs. A research task reads the web. A scheduling task reads a calendar. They should not share a toolbox.
- Write one instruction file per specialist. State the job in two or three sentences, name the allowed tools, and describe what a finished result looks like.
- Assign reasoning effort per agent. Fact checking and math get high effort. Formatting and sorting get low. This single choice often cuts cost the most.
- Add a dispatcher. Its only job is to read the incoming request, choose one specialist and pass along the needed context. Keep its instructions short.
- Test each specialist alone before wiring them together. Give it a real task and check the output against your description of done.
- Run a full request through the dispatcher and watch where it stalls or overspends. Tighten the tool list on whichever agent is heaviest.
What a good result looks like
A request arrives, the dispatcher routes it in one step, and the right specialist answers without touching unrelated tools. Costs drop because small agents carry small context. Answers arrive faster because no worker is weighing options it will never use. You can point to any agent and say exactly what it may do, which makes review and debugging far easier than untangling one giant assistant.
Build a team of narrow sub-agents instead of one overloaded assistant, so each task runs with only the tools and effort it needs. For anyone managing AI cost and speed.
Read this page, then check whether AI assistants can read yours: one scan, six pillars, a score.