AG2: Open-Source AgentOS for AI Agents

0.11.5 · active · verified Mon Apr 13

AG2 (formerly AutoGen) is an open-source programming framework for building AI agents and facilitating cooperation among multiple agents to solve tasks. It provides fundamental building blocks to create, deploy, and manage AI agents, supporting various LLMs, tool use, autonomous and human-in-the-loop workflows, and multi-agent conversation patterns. The current version is 0.11.5 and it maintains a rapid release cadence with frequent updates and new features.

Warnings

breaking AG2 is transitioning from its original framework (autogen.agentchat) to a new, redesigned beta framework (autogen.beta) which will become the official v1.0. This will involve deprecations and architectural changes in upcoming minor versions (v0.12, v0.13, v0.14) before the beta becomes stable at v1.0. Users should plan for migration.
Fix: Monitor the official AG2 documentation and release roadmap for migration guides and adopt the `autogen.beta` API for new projects to ensure future compatibility.
deprecated The `GPTAssistantAgent` class is deprecated as of v0.12 and will be removed in v0.14. Similarly, the `Swarm` orchestration pattern (and related functions like `initiate_swarm_chat()`) has been deprecated since v0.9 in favor of the new `GroupChat` pattern.
Fix: Migrate from `GPTAssistantAgent` to `ConversableAgent`. For multi-agent orchestration, use the unified `GroupChat` pattern and its associated classes (`GroupChat`, `GroupChatManager`).
gotcha LLM provider packages are not installed by default with `pip install ag2`. Users must explicitly install them as extras (e.g., `pip install "ag2[openai]"`) for their chosen LLM provider to function.
Fix: Install AG2 with the appropriate extra for your desired LLM provider, e.g., `pip install "ag2[openai]"`, `"ag2[gemini]"`, `"ag2[anthropic]"`, etc.
gotcha The `LLMConfig` object uses `deepcopy` internally to prevent unintended modifications. If `llm_config` contains custom objects that do not implement a `__deepcopy__` method, it can lead to `TypeError`.
Fix: Ensure any custom objects passed within `LLMConfig` implement the `__deepcopy__` method to support deep copying.
security As of v0.11.4, `ShellExecutor` now uses `shell=False` with `shlex.split` to prevent shell command injection vulnerabilities. Previously, users might have inadvertently created insecure execution environments.
Fix: Review any custom shell command execution logic. Ensure that commands passed to `ShellExecutor` are properly sanitized and do not rely on `shell=True` behavior if not explicitly intended and secured.

Install

pip install "ag2[openai]" Install with OpenAI support
pip install "ag2[gemini,anthropic]" Install with multiple LLM providers
pip install ag2 Minimal installation (no LLM providers)

Imports

ConversableAgent
```
from autogen import ConversableAgent
```
The primary class for creating flexible, conversational AI agents.
LLMConfig
```
from autogen import LLMConfig
```
Used to define the configuration for Language Model interactions.
GroupChat
```
from autogen import GroupChat
```
For orchestrating conversations among multiple agents.
GroupChatManager
```
from autogen import GroupChatManager
```
Manages the flow and speaker selection within a GroupChat.
Agent (beta)
```
from autogen.beta import Agent
```
The `autogen.beta` module introduces a redesigned agent framework, which will become the official `ag2` v1.0 API. While `from ag2 import Agent` might be seen in some older examples or specific contexts, the future-proof path is `autogen.beta.Agent`.

Quickstart

This quickstart initializes a `ConversableAgent` with an OpenAI LLM configuration and runs a single-turn conversation. It demonstrates basic agent creation and interaction, requiring the `OPENAI_API_KEY` environment variable to be set.

import os
from autogen import ConversableAgent, LLMConfig

# Ensure your OpenAI API key is set as an environment variable
# For example: export OPENAI_API_KEY="YOUR_API_KEY"

openai_api_key = os.environ.get("OPENAI_API_KEY", "")

if not openai_api_key:
    print("Error: OPENAI_API_KEY environment variable is not set.")
    exit()

llm_config = LLMConfig(
    {
        "api_type": "openai",
        "model": "gpt-5-nano",
        "api_key": openai_api_key
    }
)

# Create a poetic AI assistant
my_agent = ConversableAgent(
    name="helpful_agent",
    system_message="You are a poetic AI assistant, respond in rhyme.",
    llm_config=llm_config,
)

# Run the agent with a prompt
response = my_agent.run(
    message="In one sentence, what's the big deal about AI?",
    max_turns=1, # Limit turns for a quick, non-interactive example
    user_input=False, # Disable human input for automatic execution
)

# Print the agent's final response
if response.chat_history:
    print(response.chat_history[-1]["content"])
else:
    print("No response generated.")

view raw JSON →