README.md

Agentic Framework Interoperability Over the Network

This repository demonstrates how to orchestrate interactions between Langgraph, Autogen and Llama-Index agents over a network, rather than in a single codebase. By separating the server-side Autogen/Llama-Index agent (for generating short stories) and the client-side Langgraph agent, you gain a clear, modular architecture that showcases how distributed AI agents can interoperate via REST APIs.

With this example, you can:

• Isolate Agent Logic – Keep the story-generation functionality on the server, while the client remains focused on orchestration.
• Leverage REST Endpoints – Interact with the Autogen agent through streaming or stateless endpoints, providing flexible integration strategies for your own applications.
• Streamline Development – Easily experiment and debug each component, reducing complexity by splitting responsibilities across separate processes.

Server-Side

Server has an autogen (v0.4) and llama-index agents to create short stories.

Prereq:

Export variables or create a .env file (preferred) with OpenAI or Azure OPenAI infomation. Examples

export azure openai env vars for using autogen agent-

AZURE_OPENAI_API_KEY="<your-key>"
AZURE_OPENAI_ENDPOINT="<your-endpoint>"
AZURE_OPENAI_DEPLOYMENT="gpt-4o"
AZURE_API_VERSION="2024-08-01-preview"

OPENAI_API_KEY=sk-<your key>
OPENAI_MODEL_NAME=gpt-4o

---

Start server:

cd server\ap_server
python main

---

The call to RemoteGraph() on the client side sends a request to /runs/stream endpoint at server\ap_server\routers\stateless_runs.py

/runs endpoint calls the autogen agent without streaming

curl --location 'http://0.0.0.0:8123/runs' \
--header 'accept: application/json' \
--header 'Content-Type: application/json' \
--data '{
  "input": [
    {"query": "write a short story about a cat"}
  ]
}'

output:
{
    "query": "write a short story about a cat",
    "output": "<full story here>"
}

/runs/stream endpoint is used by remotegraph client

for autogen subgraph:
curl --location 'http://localhost:8123/runs/stream' \
--header 'Content-Type: application/json' \
--data '{
           "input": {
             "messages": [
               {
                 "content": "write a story about cat"
               }
             ]
           },
           "assistant_id": "autogen",
           "config": {
             "tags": ["tag1", "tag2"]
           }
         }'

for llama_index subgraph:
curl --location 'http://localhost:8123/runs/stream' \
--header 'Content-Type: application/json' \
--data '{
           "input": {
             "messages": [
               {
                 "content": "write a story about cat"
               }
             ]
           },
           "assistant_id": "llama_index",
           "config": {
             "tags": ["tag1", "tag2"]
           }
         }'

/runs/stream/agent calls the autogen agent with streaming

curl --location 'http://0.0.0.0:8123/runs/stream/agent' \
--header 'accept: application/json' \
--header 'Content-Type: application/json' \
--data '{
  "input": [
    {"query": "write a short story about a cat"}
  ]
}'

output:
<Streamed output from agent>

Client-Side

client\lg_rg.py has a client that consist of a Graph + a RemoteGraph() API that hits the above mentioned server.

client\lg_rg_human_in_loop_autogen.py has a client that consist of a Graph + a RemoteGraph() API that hits the above mentioned server via /runs/human_in_loop endpoint and runs a human_in_loop workflow in the autogen subgraph.

client\llama_rg.py has a client that consist of a Graph + a RemoteGraph() API that hits the above mentioned server.

client\rest.py contains a client that makes a stateless request to the above mentioned server via /runs endpoint.

client\rest_human_in_loop_autogen.py contains a client that makes a stateless request to the above mentioned server via /runs/human_in_loop endpoint and runs a human_in_loop workflow in the autogen subgraph.

client\rest_human_in_loop_lg.py contains a client that makes a stateless request to the above mentioned server via /runs endpoint and then sends the output to a langgraph for human in loop.

Testing

• Run Server as python main.py
• Run RemoteGraph client as python lg.py
• Run RemoteGraph client for human-in-loop demo for autogen python lg_human_in_loop_autogen.py
• Run stateless REST client as python rest.py
• Run stateless REST client for human-in-loop demo for autogen python rest_human_in_loop_autogen.py
• Run stateless REST client for human-in-loop demo for LangGraph python rest_human_in_loop_lg.py