Dynamic Agent Delegation: The supervisor can decide whether to handle a user query itself or delegate it to a configured specialist agent.
Configurable Agents: Easily define and configure multiple child agents, selecting from the list of existing agents you have deployed and configured in Open Agent Platform (OAP).
Customizable System Prompt: Tailor the supervisor's behavior and instructions using a configurable system prompt, with a sensible default provided.

Deployments

Local

To run the supervisor locally,

Clone this repo

gh repo clone langchain-ai/open-agent-supervisor

Copy the .env.example file into .env, and set your environment variables

We default to OpenAI to provide the underlying supervisor agent LLM

cp .env.example .env

Install the in-memory LangGraph command line package

pip install -U "langgraph-cli[inmem]"

Run the LangGraph server locally

langgraph dev

For more info, see our LangGraph CLI docs

OAP

Important

Prerequisites: Have at least one supervisor deployed to your instance of LangGraph Platform (LGP) that matches or extends the base implementation provided in this repo

To add a supervisor agent to the platform:

Navigate to the Agents tab in the left menu bar
Create and agent and select the graph deployed to your LGP instance described in the prerequisites above
Give your supervisor a name, description, an optional system prompt if you'd like to modify it and select the agents that it will orchestrate (note - these agents will need to already be deployed in your LGP instance and configured in Open Agent Platform)

How it Works

The supervisor operates based on a system prompt that instructs it on how to manage incoming user messages. It can either:

Answer the user directly.
Delegate the query to a specialist agent by invoking a tool in the format delegate_to_<agent_name>(user_query).

The user sees all messages and and optionally all tool calls, ensuring transparency in the conversation flow.

Authentication

This project uses LangGraph custom auth to authenticate requests to the server. It's configured to use Supabase as the authentication provider, however it can be easily swapped for another service.

Requests must contain an Authorization header with a Bearer token. This token should be a valid JWT token from Supabase.

The auth handler then takes that token and verifies it with Supabase. If the token is valid, it returns the user's identity. If the token is invalid, it raises an exception. This means you must have a Supabase URL & key set in your environment variables to use this auth handler:

SUPABASE_URL=""
SUPABASE_KEY=""

The auth handler is then used as middleware for all requests to the server. It is configured to run on the following events:

threads.create
threads.read
threads.delete
threads.update
threads.search
assistants.create
assistants.read
assistants.delete
assistants.update
assistants.search
store

For creation methods, it auto-injects the user's ID into the metadata. This is then uses in all read/update/delete/search methods to ensure that the user can only access their own threads and assistants.

By using custom authentication, we can call this LangGraph server directly from a frontend application, without having to worry about exposing API keys/secrets, since you only need a JWT token from Supabase to authenticate.

For more info, see our LangGraph custom auth docs.