This example runs a gRPC server using WorkerAgentRuntimeHost and instantiates three distributed runtimes using WorkerAgentRuntime. These runtimes connect to the gRPC server as hosts and facilitate a round-robin distributed group chat. This example leverages the Azure OpenAI Service to implement writer and editor LLM agents. Agents are instructed to provide concise answers, as the primary goal of this example is to showcase the distributed runtime rather than the quality of agent responses.

Setup

Setup Python Environment

Create a virtual environment as instructed in README.
Run uv pip install chainlit in the same virtual environment

General Configuration

In the config.yaml file, you can configure the client_config section to connect the code to the Azure OpenAI Service.

Authentication

The recommended method for authentication is through Azure Active Directory (AAD), as explained in Model Clients - Azure AI. This example works with both the AAD approach (recommended) and by providing the api_key in the config.yaml file.

Run

Run Through Scripts

The run.sh file provides commands to run the host and agents using tmux. The steps for this approach are:

Install tmux.
Activate the Python environment: source .venv/bin/activate.
Run the bash script: ./run.sh.

Here is a screen recording of the execution:

Note: Some asyncio.sleep commands have been added to the example code to make the ./run.sh execution look sequential and visually easy to follow. In practice, these lines are not necessary.

Run Individual Files

If you prefer to run Python files individually, follow these steps. Note that each step must be run in a different terminal process, and the virtual environment should be activated using source .venv/bin/activate.

python run_host.py: Starts the host and listens for agent connections.
python run_editor.py: Starts the editor agent and connects it to the host.
python run_writer.py: Starts the writer agent and connects it to the host.
chainlit run run_group_chat_manager.py --port 8001: Run chainlit app which starts group chat manager agent and sends the initial message to start the conversation. We're using port 8001 as the default port 8000 is used to run host (assuming using same machine to run all of the agents)

What's Going On?

The general flow of this example is as follows:

The Group Chat Manager, on behalf of User, sends a RequestToSpeak request to the writer_agent.
The writer_agent writes a short sentence into the group chat topic.
The editor_agent receives the message in the group chat topic and updates its memory.
The Group Chat Manager receives the message sent by the writer into the group chat simultaneously and sends the next participant, the editor_agent, a RequestToSpeak message.
The editor_agent sends its feedback to the group chat topic.
The writer_agent receives the feedback and updates its memory.
The Group Chat Manager receives the message simultaneously and repeats the loop from step 1.

Here is an illustration of the system developed in this example:

graph TD;
    subgraph Host
        A1[GRPC Server]
        wt[Writer Topic]
        et[Editor Topic]
        gct[Group Chat Topic]
    end

    subgraph Distributed Writer Runtime
        writer_agent[<img src="./public/avatars/writer.png" width="50"/> Writer Agent] --> A1
        wt -.->|2 - Subscription| writer_agent
        gct -.->|4 - Subscription| writer_agent
        writer_agent -.->|3 - Publish: Group Chat Message| gct
    end

    subgraph Distributed Editor Runtime
        editor_agent[<img src="./public/avatars/editor.png" width="50"/> Editor Agent] --> A1
        et -.->|6 - Subscription| editor_agent
        gct -.->|4 - Subscription| editor_agent
        editor_agent -.->|7 - Publish: Group Chat Message| gct
    end

    subgraph Distributed Group Chat Manager Runtime
        group_chat_manager[<img src="./public/avatars/group_chat_manager.png" width="50"/> Group Chat Manager Agent] --> A1
        gct -.->|4 - Subscription| group_chat_manager
        group_chat_manager -.->|1 - Request To Speak| wt
        group_chat_manager -.->|5 - Request To Speak| et
    end

    style wt fill:#beb2c3,color:#000
    style et fill:#beb2c3,color:#000
    style gct fill:#beb2c3,color:#000
    style writer_agent fill:#b7c4d7,color:#000
    style editor_agent fill:#b7c4d7,color:#000
    style group_chat_manager fill:#b7c4d7,color:#000

TODO:

Properly handle chat restarts. It complains about group chat manager being already registered
Send Chainlit messages within each agent (Currently the manager can just sends messages in the group chat topic)
Add streaming to the UI like this example but Autogen's Open AI Client does not supporting streaming yet