Maybe in some months time when you’re reading this, the fad would be over. But for now, open agent frameworks in Python are essential for building autonomous AI systems that can reason, plan, and act, particularly in the context of large language models (LLMs). These frameworks enable developers to create agents for tasks ranging from conversational AI to complex multi-agent workflows. However, it’s not all good news so we feel a bit spoiled for choice, since there’s a plethora of the so-called agentic frameworks, an the number of them more keep growing more and more.
Source: image generated with gpt4o
What's an "agentic framework"?
An agentic framework provides the architecture to build software agents that can reason, plan, and act autonomously. These frameworks offer abstractions and orchestration logic to manage prompts, memory, tools, and interactions between agents.
Among the options we find some nice tech pieces:
- LangChain / LangGraph: LangChain is a leading framework with over 108,000 GitHub stars, offering extensive integrations and tools for building AI applications, including an “Agents” module. LangGraph, part of the ecosystem, provides low-level orchestration for more granular control. It’s trusted by enterprises like Replit, Rakuten, and Klarna, with over 1 million practitioners using it.
- CrewAI: With 31,800 GitHub stars, CrewAI is a lightweight, fast framework for multi-agent orchestration, independent of LangChain. It’s used by 60% of Fortune 500 companies and in 150+ countries, offering both high-level simplicity (Crews) and precise control (Flows).
- Microsoft AutoGen: Developed by Microsoft, AutoGen has 44,700 GitHub stars and focuses on multi-agent collaboration with dynamic workflows. It supports conversational interfaces and is used in research and production for complex AI workflows.
- OpenAI Agents SDK: With 10,400 GitHub stars, this lightweight framework from OpenAI supports multi-agent workflows, is provider-agnostic, and works with over 100 LLMs. It’s ideal for rapid prototyping and scalable AI systems.
- Agno: Formerly Phidata, Agno is gaining traction for its high performance, with fast agent instantiation (3μs) and low memory use (6.5Kib). It supports multi-modality (text, image, audio, video) and is used for advanced, memory-intensive tasks.
- Llamaindex: With over 4 million monthly downloads, Llamaindex specializes in context-augmented agents, particularly for document processing (handling 300+ formats). It’s trusted by enterprises like KPMG, Salesforce, and Rakuten, with 1.5k+ contributors and 200M+ pages processed.
LangChain
Maybe the first serious player in the open source area, but in my view over-complicated and too much engineerish (while other oriented agent definition to organizational and mission aspects). In any case, tt has become the flagship framework for building LLM-powered applications, with LangGraph offering low-level orchestration. It provides a full suite including LangSmith and LangGraph Platform, ensuring seamless integration for development, deployment, and monitoring.
- Key Features:
- Wide range of integrations with LLMs (e.g., OpenAI, Hugging Face).
- Agents module for creating and testing autonomous agents.
- Support for memory, tools, and human-in-the-loop workflows.
- LangGraph enables durable execution, debugging, and memory management.
- Traction: Over 108,000 GitHub stars for LangChain, with over 1 million practitioners using the frameworks. Trusted by enterprises like Replit LangChain Customer Replit, Rakuten LangChain Customer Rakuten, and Klarna LangChain Customer Klarna.
- Use Cases: Ideal for conversational AI, task-oriented agents (e.g., web search, database queries), and RAG systems. It’s particularly popular for general-purpose agent development due to its extensive ecosystem.
- Why It’s Popular: The largest developer community in GenAI, with comprehensive tools and strong documentation, makes it a go-to choice for developers.
Info: LangChain vs. LangGraph
While LangChain offers higher-level APIs and integrations, LangGraph is a lower-level orchestration tool designed for fine-grained control. They work together, but they cater to slightly different developer needs.
Llamaindex
Another classical contender, although it orientation is slightly different to Langchain, it eventually has all you need to create an orchestrated agent app. Its speciality is context-augmented AI agents, with strong features in document processing and retrieval-augmented generation (RAG). It’s loved by developers and trusted by enterprises for its document handling capabilities (perhaps due to its greater simplicity compared to langchain).
- Key Features:
- World’s best document processing layer, handling 300+ formats (PDFs, PowerPoints, spreadsheets) via LlamaCloud LlamaCloud Sign Up.
- Supports building agents with context augmentation, leveraging extensive connectors and tools via LlamaHub.
- Extensive documentation, including building blocks, workflows, and notebooks, with a strong developer network LlamaIndex Discord.
- Traction: Over 4 million monthly downloads, 1.5k+ contributors, and 200M+ pages processed. Trusted by enterprises like KPMG, Salesforce, Rakuten, and Cemex LlamaIndex Official Website.
- Use Cases: Powers agentic document workflows in finance, insurance, manufacturing, retail, and technology. Ideal for RAG systems and knowledge-intensive applications.
- Why It’s Popular: Its focus on document processing and enterprise-grade reliability makes it a top choice for context-augmented agents.
RAG ≠ Agents, but they can be close friends
Retrieval-Augmented Generation (RAG) enhances LLM responses by injecting relevant external data. Frameworks like LlamaIndex bridge RAG techniques with agent-based logic for smarter outputs.
Crewai
They propose a lean, lightning-fast Python framework built from scratch, independent of LangChain. It focuses on multi-agent orchestration, offering both high-level simplicity (Crews) and precise, event-driven control (Flows). All these aspects, in my view, make this framework more accessible and creates a more maintainable for agent goals definition.
- Key Features:
- High-level abstraction for role-based agent creation (Crews).
- Granular control over workflows (Flows) for tailored, collaborative intelligence.
- Supports multi-agent collaboration and communication, with real-time management dashboards.
- Lightweight design, suitable for cloud, self-hosted, or local deployment.
- Traction: Over 31,800 GitHub stars, used by 60% of Fortune 500 companies and in 150+ countries. Offers a complete suite for building, deploying, and managing AI agents, with over 100,000 developers certified through community courses CrewAI Learn.
- Use Cases: Building multi-agent systems for complex tasks like data research, automation, and content generation. It’s ideal for role-based agents in industries like healthcare, finance, and marketing .
- Why It’s Popular: Its enterprise-ready features, ease of use, and strong adoption by large organizations make it a top choice for multi-agent systems.
OpenAI Agents SDK
The answer of OpenAI to agent matter is lightweight framework for building multi-agent workflows, provider-agnostic, and supporting over 100 LLMs beyond OpenAI’s. It emphasizes simplicity and flexibility for rapid prototyping and scalable systems.
- Key Features:
- Lightweight design for easy integration into existing workflows.
- Supports agents, handoffs, guardrails, and tracing for debugging and optimization.
- Compatible with various LLMs, ensuring flexibility in model choice.
- Built-in tracing integrations with tools like Logfire, AgentOps, and Braintrust OpenAI Agents Tracing.
- Traction: Over 10,400 GitHub stars, directly from OpenAI, ensuring alignment with cutting-edge AI developments. Gaining adoption for its simplicity and flexibility .
- Use Cases: Building multi-agent systems for task delegation, creating agents that dynamically switch between models or tools, and developing scalable AI workflows with real-time monitoring.
- Why It’s Popular: Its provider-agnostic approach and lightweight design make it accessible for developers looking for flexibility.
Multi-agent ≠ multi-model
A multi-agent system means different agents performing specialized tasks collaboratively. It doesn’t necessarily imply using different LLMs—though frameworks like Agno and OpenAI SDK make that possible too.
Microsoft AutoGen
Another big fish proposition, it has been developed by Microsoft, is an open-source framework for building multi-agent AI systems that can collaborate and solve tasks autonomously. It includes components like AutoGen Studio, AgentChat, Core, and Extensions for scalable applications.
- Key Features:
- Multi-agent collaboration with conversational interfaces.
- Dynamic workflows, natural language interactions, and human-in-the-loop support.
- Built-in tools for task automation, optimization, and integration with LLMs.
- Offers a visual interface for prototyping (AutoGen Studio) and advanced multi-agent systems (Core).
- Traction: Over 44,700 GitHub stars, backed by Microsoft, ensuring enterprise-grade reliability. Used in research and production environments, with a focus on scalability .
- Use Cases: Building conversational AI systems, automating complex workflows (e.g., code generation, data analysis), and facilitating human-AI collaboration in research and development.
- Why It’s Popular: Its robust support for multi-agent systems and Microsoft’s backing provide long-term maintenance and enterprise trust.
Agno
A recent incorporation to the area isAgno, formerly Phidata, is a high-performance, lightweight framework for building multi-agent systems with memory, knowledge, and reasoning. It’s designed for speed, flexibility, and multi-modality, supporting text, image, audio, and video.
- Key Features:
- Model-agnostic, supporting 23+ model providers (e.g., OpenAI, Anthropic, Mistral) with no lock-in.
- Extremely fast agent instantiation (3μs) and low memory usage (6.5Kib), tested on Apple M4 MacBook Pro.
- Natively multi-modal, accepting and generating diverse input/output types.
- Built-in support for reasoning (3 approaches: Reasoning Models, ReasoningTools, chain-of-thought), memory, and multi-agent collaboration.
- Pre-built FastAPI routes for rapid deployment to production, with real-time monitoring on app.agno.com.
- Traction: Gaining traction for its performance, with growing community interest. Used by developers for advanced, multi-modal agents, with documentation at Agno Docs.
- Use Cases: Building multi-modal agents for creative applications (e.g., image generation, video analysis), creating agents with persistent memory and knowledge bases, and developing teams of specialized agents for complex workflows.
- Why It’s Popular: Its unmatched speed and efficiency, especially compared to frameworks like LangGraph (claimed ~10,000x faster), make it ideal for performance-critical applications.
Hugginface’s smolagents
The proposal from HuggingFace is truly gaining attention for its simple, code-driven approach. It allows agents to write actions as Python code snippets, making it efficient for certain tasks. Launched in late 2024, smolagents is still growing but is already popular for its lightweight design. It’s great for developers who want to build agents for tasks like data research, automation, and content generation, especially where code execution is key. However, it may not yet have the same level of community support as older frameworks like LangChain, but prospects are promising.
- Key Features:
- Simplicity: Entire agent logic in ~1,000 lines of code, making it easy to inspect and understand.
- Code Agents: Agents write actions in Python code for better composability and efficiency, reducing steps and LLM calls by about 30% compared to standard tool-calling methods.
- Security: Supports sandboxed code execution via E2B or Docker, with options like a secure Python interpreter for safe execution.
- Model-agnostic: Compatible with various LLMs from Hugging Face, OpenAI, Anthropic, etc., and integrates seamlessly with the Transformers library.
- Multi-modal support: Handles text, vision, video, and audio inputs, similar to advanced frameworks like Agno.
- Tool compatibility: Integrates with tools from LangChain, Anthropic’s MCP, or Hugging Face Spaces, and allows easy definition of custom tools.
- Hub integrations: Seamlessly share and load agents and tools to/from the Hugging Face Hub as Gradio Spaces, fostering collaboration.
- Traction: Launched in late 2024, smolagents is gaining traction for its minimalist approach. While specific GitHub stars aren’t mentioned, it’s being adopted and discussed in various blogs and tutorials, such as demos for getting stock prices or upvoted papers. It’s recognized as a distinct framework with strengths in code-driven patterns, with 152 contributors noted on its GitHub page .
- Use Cases: Ideal for building agents that require code execution, such as data research, automation, and content generation. Suitable for developers who want a lightweight yet powerful framework for agentic workflows, especially for tasks like fetching data or executing complex computations.
- Why It’s Popular: Its minimalist design and focus on code-driven agents make it appealing for developers who value transparency and efficiency, particularly for rapid prototyping and small-scale projects. Its integration with Hugging Face’s ecosystem also enhances its adoption.
Summary table
To provide a clear overview, the following table compares the frameworks based on key metrics and strengths (stars as of 1st of July 2025):
| Framework | GitHub Stars | Key Strengths | Use Cases |
|---|---|---|---|
| LangChain | 108k+ | Comprehensive tools, large community, integrations | Conversational agents, RAG, task automation |
| CrewAI | 31.8k | Lightweight, multi-agent orchestration, enterprise | Role-based agents, complex workflows |
| Microsoft AutoGen | 44.7k | Multi-agent collaboration, dynamic workflows | Conversational AI, task automation |
| OpenAI Agents SDK | 10.4k | Lightweight, provider-agnostic, tracing | Multi-agent workflows, task delegation |
| Agno | 29.1k | High performance, multi-modality, reasoning | Multi-modal agents, memory-intensive tasks |
| Llamaindex | 42.7k | Document processing, RAG, enterprise-ready | Context-augmented agents, knowledge-intensive |
| Smolagents | 20.8k | Minimalist, code-driven agents, multi-modal | Code execution agents, lightweight development |
This table highlights the diversity in strengths, with smolagents standing out for its minimalist, code-driven approach, while established frameworks like LangChain lead in community size and enterprise adoption.
Why GitHub stars matter (and don’t)
While stars reflect community interest and maturity, they don’t always signal the best fit for your use case. Emerging tools like Agno may outperform larger frameworks in specific scenarios.
Final remarks
The most used open agent frameworks in Python, based on community traction, features, and real-world adoption, include LangChain/LangGraph, CrewAI, Microsoft AutoGen, OpenAI Agents SDK, Agno, Llamaindex, and the newly introduced smolagents. Each framework has unique strengths, and the choice depends on the specific use case. In my view, the decision will depends on the type of application, a more agent orchestration app would go for agno, crewai or autogen while a more information-eager one would take Llamaindex. On the other hand, if you need to tightlty control the workflow Langhcain will be your tool.