One agent can't be an expert at everything — and a single huge system prompt gets brittle. Sub-agents let you compose a team of focused specialists behind one entry point, coordinated by a supervisor that routes each request to the right expert.
A sub-agent is just another saved agent that a parent agent is allowed to delegate to. You add sub-agents in step 5 of the builder. Each sub-agent keeps its own model, system prompt, knowledge, and tools — so a research specialist can run a web-search-heavy model while a writer runs a different one, all under the same parent.
When an agent has sub-agents, it becomes a supervisor. Instead of answering directly, it interprets the request, decides which specialist (or specialists) should handle each part, delegates the work, and then merges the results into one coherent answer. The person asking sees a single agent; behind the scenes a coordinated team did the work.
| Benefit | Why it matters |
|---|---|
| Focus | Each sub-agent has a tight system prompt and only the tools it needs, so it performs its one job better than a generalist. |
| Right model for the job | A reasoning-heavy researcher and a fast drafter can run on different models from your six configured providers — chosen per sub-agent. |
| Reuse | A sub-agent is a normal agent, so the same specialist can be shared across many supervisors instead of duplicated. |
| Maintainability | Improving one specialist improves every team it belongs to, without touching the others. |
Imagine a Content Desk supervisor for your marketing team. You give it three sub-agents:
A teammate asks the Content Desk for a launch announcement. The supervisor sends the brief to the researcher, hands the findings to the writer, routes the draft through the reviewer, and returns a finished, on-brand announcement — all from one request.
The supervisor pattern executes on the same LangGraph runtime as every other agent and workflow. The supervisor and its sub-agents are coordinated as a graph, so routing, delegation, and merging all happen inside one managed run with shared budgets and event streaming.
Give the supervisor a clear routing prompt — when to use each sub-agent and how to combine their output — and let the specialists hold the deep instructions. A thin supervisor over focused experts is easier to reason about and to improve over time.