A small demo where a supervisor deep agent fans out parallel async researcher subagents (each backed by a separate LangGraph graph), streams updates in a Vite + React UI, and polls run status until tasks finish.

Prerequisites

Node.js and pnpm
From the monorepo root: pnpm install (and pnpm build if deepagents has not been built yet)

Environment

Copy the example env file and fill in keys:

cp examples/async-subagents/parallel-research/.env.example examples/async-subagents/parallel-research/.env

Set ANTHROPIC_API_KEY (used by the models).
Optionally set TAVILY_API_KEY for real web search. Without it, the researcher uses a stub search implementation so the flow still runs.

1. Run the LangGraph dev server

From the repository root:

npx @langchain/langgraph-cli dev -c examples/async-subagents/parallel-research/langgraph.json

This loads langgraph.json, which registers two graphs:

supervisor — orchestrates async researcher tasks (this is what the UI talks to).
researcher — research subagent graph invoked asynchronously by the supervisor.

Leave this process running.

2. Run the web UI

In a second terminal, from the repository root:

pnpm --filter @examples/async-subagents-parallel-research dev

Or from the UI package directory:

cd examples/async-subagents/parallel-research/ui && pnpm dev

Open the URL Vite prints (usually http://localhost:5173).

Using the demo

Normal questions are answered by the supervisor directly.
To trigger parallel researchers, use wording that clearly asks for research, e.g. “Research …”, “Investigate …”, “Deep dive into …” (see the supervisor system prompt in supervisor.ts for the full trigger behavior).
The UI shows researcher cards for async tasks and can fetch results when a run completes.

Project layout

Path	Role
`supervisor.ts`	Supervisor graph (`assistantId: "supervisor"` in the UI)
`researcher.ts`	Researcher graph (`graphId: "researcher"`)
`langgraph.json`	LangGraph CLI config for both graphs
`ui/`	Vite + React client using `@langchain/langgraph-sdk` streaming