Why agentic systems went mainstream
In 2024, agents were experimental. In 2026, it's a production pattern: GitHub Copilot Workspace, Cursor, Devin — all run on agentic loops. On GitHub, 50k+ repositories with the "ai-agent" tag were added in the last year.
AutoGen: agents as dialogue
Microsoft AutoGen 0.4 rewritten on event-driven architecture. Agents communicate via messages, each is a separate process. Strong point: GroupChat lets you launch a team of agents where each has a role.
pip install autogen-agentchat
import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.teams import RoundRobinGroupChat
from autogen_ext.models.openai import OpenAIChatCompletionClient
model_client = OpenAIChatCompletionClient(model="gpt-4o")
agent1 = AssistantAgent("researcher", model_client=model_client,
system_message="Ты исследователь. Найди факты.")
agent2 = AssistantAgent("writer", model_client=model_client,
system_message="Ты автор. Оформи результат.")
team = RoundRobinGroupChat([agent1, agent2], max_turns=4)
async def main():
result = await team.run(task="Сделай обзор Llama 3")
print(result.messages[-1].content)
asyncio.run(main())
CrewAI: roles and tasks
CrewAI emphasizes declarativity: you describe agents as employees with roles, tasks as tickets. 30k+ stars, the fastest-growing agentic framework of 2025.
from crewai import Agent, Task, Crew
researcher = Agent(
role='AI Researcher',
goal='Найти актуальные данные по теме',
backstory='Эксперт по анализу технической документации',
verbose=True
)
writer = Agent(
role='Technical Writer',
goal='Написать понятный технический текст',
backstory='Пишет для разработчиков 10 лет',
)
task = Task(description='Напиши статью о LangChain vs LlamaIndex',
agent=researcher)
crew = Crew(agents=[researcher, writer], tasks=[task])
result = crew.kickoff()
LangGraph: graph as code
LangGraph — part of the LangChain ecosystem. Agent = state graph. Full control: loops, conditional transitions, human-in-the-loop pauses. More complex, but more predictable in production.
Final comparison
| Criterion | AutoGen | CrewAI | LangGraph |
|---|---|---|---|
| Entry threshold | medium | low | high |
| Flow control | medium | low | full |
| Multi-agent capability | excellent | good | good |
| Production readiness | growing | good | excellent |
FAQ
Can I use local models via Ollama?
Yes, all three frameworks support OpenAI-compatible endpoints. Change base_url to localhost:11434/v1 and model to llama3.
What's better for code automation?
AutoGen with CodeExecutorAgent — specifically optimized for writing and executing code. Built-in sandbox isolation.