AutoGen / AG2

The classic autogen package (ConversableAgent, LLMConfig). The redesigned ag2 1.x API has no ConversableAgent; this page doesn't apply to it.

Connect

BASE is http://127.0.0.1:4010/v1 and KEY is your plugboard key. Install autogen[openai,anthropic].

BASE = os.environ.get("PLUGBOARD_BASE_URL", "http://127.0.0.1:4010/v1")
KEY = os.environ["PLUGBOARD_KEY"]


def llm_config(model: str = "fast", stream: bool = False) -> LLMConfig:
    return LLMConfig(
        {
            "api_type": "openai",
            "model": model,
            "base_url": BASE,
            "api_key": KEY,
            "max_tokens": 1024,
            "stream": stream,
        }
    )

bot = ConversableAgent("bot", llm_config=cfg, human_input_mode="NEVER")
reply = bot.generate_reply(messages=[{"role": "user", "content": "Reply with exactly: pong"}])

Pass the endpoint in the entry dict; keyword arguments to LLMConfig don't set it. llm_config(stream=True) streams. Here cfg is llm_config().

A tool round trip is two agents: the assistant registers the function for the model, the user agent executes it.

assistant = ConversableAgent(
    "assistant",
    llm_config=cfg,
    human_input_mode="NEVER",
    system_message="Answer weather questions by calling get_weather.",
)
user = ConversableAgent(
    "user",
    llm_config=False,
    human_input_mode="NEVER",
    default_auto_reply="",
    is_termination_msg=lambda m: False,
)
assistant.register_for_llm(description="Get the current weather for a city.")(get_weather)
user.register_for_execution()(get_weather)
result = user.initiate_chat(
    assistant,
    message="What is the weather in Paris? Use the tool.",
    max_turns=4,
    silent=True,
)

Anthropic route

With api_type: "anthropic" AutoGen calls the Messages API. Its Anthropic client adds /v1/messages itself, so base_url is the plugboard address without /v1. Chat and the tool round trip above run unchanged with cfg = anthropic_config().

def anthropic_config(model: str = "fast") -> LLMConfig:
    return LLMConfig(
        {
            "api_type": "anthropic",
            "model": model,
            # The Anthropic SDK adds /v1/messages itself: pass the address without /v1.
            "base_url": BASE.removesuffix("/v1"),
            "api_key": KEY,
            "max_tokens": 1024,
        }
    )

The code is in examples/autogen/:

cd examples/autogen
PLUGBOARD_KEY=<app key> uv run chat.py

What works

Chat Completions and Messages, each with chat and function tools (register_for_llm and register_for_execution). Chat Completions also streams.

Limits

• autogen logs Model fast is not found. The cost will be 0. for an alias. Add a price field to the entry to silence it.
• A tool round takes four turns (prompt, call, result, answer). Set max_turns to at least 4 and don't terminate on the first message.
• The Anthropic client of AutoGen doesn't stream.