AutoGen

0.10.0 · active · verified Wed Apr 15

AutoGen is a framework that allows for the development of LLM applications using multiple agents that can converse with each other to solve tasks. The `pyautogen` PyPI package (current version 0.10.0) acts as a proxy, primarily installing `autogen-agentchat` to provide the core multi-agent conversation capabilities. It has a relatively rapid release cadence, with minor versions often released monthly or more frequently, introducing new features and breaking changes.

Warnings

gotcha The PyPI package is `pyautogen`, but the modules are imported from the `autogen` namespace (e.g., `import autogen`). Importing from `pyautogen` will fail.
Fix: Always use `import autogen` and `from autogen import ...` for all modules and classes.
breaking The structure of `llm_config` and `config_list` has changed multiple times between minor versions (e.g., 0.1.x, 0.2.x, 0.6.x, 0.9.x). Ensure your configuration matches the exact version's documentation.
Fix: Consult the official AutoGen documentation or GitHub README for the specific version you are using to verify the `llm_config` and `config_list` format.
breaking The `autogen.oai.Completion` module and related utilities for direct OpenAI API calls were moved or refactored. In newer versions, the primary way to interact with models is through agent `llm_config`.
Fix: Prefer configuring models via the `llm_config` dictionary when initializing agents. Avoid using `autogen.oai.Completion` directly unless specifically following an advanced pattern in recent documentation.
gotcha `UserProxyAgent`'s `human_input_mode` defaults to `ALWAYS` or `TERMINATE` depending on the version, which can pause execution indefinitely waiting for user input. For fully autonomous execution, set `human_input_mode="NEVER"`.
Fix: Explicitly set `human_input_mode="NEVER"` (for no human input), `"TERMINATE"` (for human input only at termination), or `"ALWAYS"` (for human input at every turn) during `UserProxyAgent` initialization.
gotcha API key management requires a `config_list` where models and keys are defined. While environment variables (e.g., `OPENAI_API_KEY`) are common, they must still be referenced correctly within `config_list` or an `OAI_CONFIG_LIST` JSON file.
Fix: Ensure `config_list` explicitly includes `api_key` values, either directly or by referencing environment variables (`os.environ.get("KEY")`) or by ensuring `OAI_CONFIG_LIST` file is correctly formatted and discoverable.

Install

pip install pyautogen Install pyautogen
pip install pyautogen[blend] Install with blend for local LLM support

Imports

UserProxyAgent
wrong:
```
from pyautogen import UserProxyAgent
```
correct:
```
from autogen import UserProxyAgent
```
`pyautogen` is the PyPI package, but the actual modules are under the `autogen` namespace.
AssistantAgent
wrong:
```
from pyautogen import AssistantAgent
```
correct:
```
from autogen import AssistantAgent
```
`pyautogen` is the PyPI package, but the actual modules are under the `autogen` namespace.

config_list_from_json

from autogen import config_list_from_json

GroupChat
```
from autogen.groupchat import GroupChat
```

GroupChatManager

from autogen.groupchat import GroupChatManager

Quickstart

This quickstart demonstrates setting up two basic agents: an `AssistantAgent` powered by GPT-4o and a `UserProxyAgent` to initiate the conversation and manage execution. It shows how to configure LLM settings using an environment variable for API keys and sets up basic code execution and termination conditions. Remember to set the `OPENAI_API_KEY` environment variable.

import autogen
import os

# Configure API key from environment variable or direct config_list
config_list = [
    {
        "model": "gpt-4o",
        "api_key": os.environ.get("OPENAI_API_KEY", None)
    }
]

# Define agents
assistant = autogen.AssistantAgent(
    name="Assistant",
    llm_config={
        "config_list": config_list,
        "temperature": 0.7
    }
)

user_proxy = autogen.UserProxyAgent(
    name="UserProxy",
    human_input_mode="NEVER", # Set to "ALWAYS" or "TERMINATE" for human interaction
    max_consecutive_auto_reply=10,
    is_termination_msg=lambda x: x.get("content", "").rstrip().endswith("TERMINATE"),
    code_execution_config={
        "work_dir": "coding", # Create a 'coding' directory for code execution
        "use_docker": False # Set to True to use Docker for safer execution
    }
)

# Start a chat
user_proxy.initiate_chat(
    assistant,
    message="What is the capital of France?"
)

view raw JSON →