AI Agent Frameworks
What this article covers
- Which AI agent frameworks exist and what they’re suited for.
- Which framework makes sense for which use case.
- Which frameworks run locally and which require cloud APIs.
- Links to all framework articles with installation and examples.
Why this AI framework guide?
New AI agent frameworks appear constantly, making it hard to keep track of which ones you actually need or should try. I test frameworks daily for content creation, image and video production, and coding tasks.
Whether you want to work with agents locally or in the cloud, you’ll find practical guides, installation instructions, and comparisons for the most common frameworks in today’s AI agent landscape.
Why do you need a framework?
A language model alone can only generate text. To pursue goals, use tools, and plan multiple steps, you need a framework. It handles the loop of observing, planning, and executing. Without a framework, you’d have to program this logic yourself.
Here’s an example: you want an agent to gather current news, summarize it, and write it to a file. With a framework like LangGraph or CrewAI, you define the goal and the tools. The framework manages the loop, calls the tools, and stores intermediate results. Without a framework, you’d have to code every API call, loop step, and error handling yourself.
Who is this section for?
- You want to build AI agents and are looking for the right framework.
- You’ve read the fundamentals on agents and tool calling, and now want to get practical.
- You’ll need Python basics for most frameworks. For Dify and Langflow, low-code is often enough.
- You can work locally with Ollama or use cloud APIs like OpenAI.
Framework overview
| Framework | Focus | Difficulty | Best for |
|---|---|---|---|
| LangGraph | States and decision graphs | Medium | Complex workflows |
| CrewAI | Roles and tasks | Easy | Role-based teams |
| AutoGen | Conversations and coding | Medium | Programming, debugging |
| OpenAI Agents SDK | Official SDK from OpenAI | Easy | Quick prototypes with OpenAI API |
| Dify | Visual workflow platform | Easy | Low-code agents and RAG |
| Langflow | Visual LangChain interface | Easy | Drag-and-drop prototypes |
| OpenHands | Coding agent for repositories | Medium | Code editing, refactoring |
| AgentVerse | Multi-agent simulation | Hard | Research and experiments |
| Camel AI | Role-based conversations | Medium | Negotiations and role play |
| Codebase Memory | Knowledge from code | Medium | Code knowledge for agents |
| OpenClaw | Agent orchestration | Medium | Autonomous workflows |
| Agentic AI | Enterprise agents | Easy | Business process automation |
When to choose which framework
- LangGraph if you need workflows with loops, conditions, and state management.
- CrewAI if you want to build agents with distinct roles and clearly defined tasks.
- AutoGen if you want code written, reviewed, or explained.
- OpenAI Agents SDK if you want to experiment quickly with the official SDK.
- Dify if you prefer clicking to coding and want self-hosting.
- Langflow if you want to connect LangChain components visually.
- OpenHands if you want to automate real repository changes.
- AgentVerse if you want to explore multi-agent scenarios and emergent behavior.
- Camel AI if you want to simulate role-based conversations and negotiations.
- Codebase Memory if you want long-term memory for code agents.
- OpenClaw if you want to test autonomous workflows and goal tracking.
- Agentic AI if you want to deploy AI agents in business processes.
Framework highlights
LangGraph, CrewAI, and AutoGen
These three frameworks are your entry point into local AI agents. They run with Ollama and cover most use cases:
- LangGraph - State-based workflows with decision edges.
- CrewAI - Agent teams with roles and tasks.
- AutoGen - Coding and multi-turn conversations.
Visual and low-code options
If you prefer working with an interface, Dify and Langflow offer the right tools. Both can be installed locally and connected to Ollama:
- Dify - Comprehensive platform for chatbots, RAG, and agents.
- Langflow - Visual LangChain building blocks.
Specialized and experimental frameworks
For advanced or specific use cases, these frameworks are worth exploring:
- OpenHands - Edits real repositories and executes code.
- AgentVerse - Simulates multi-agent scenarios.
- Camel AI - Role-based conversations.
- Codebase Memory - Building knowledge from code.
- OpenClaw - Open orchestration of autonomous workflows.
- Agentic AI - Enterprise approach for business processes.
Foundations before the framework
Before diving into AI agents, you should understand how tool calling works and how agents differ from chatbots.
FAQ
Which framework is best for beginners?
CrewAI is very beginner-friendly. If you prefer something more technical and state-based, start with LangGraph. For graphical interfaces, Dify or Langflow work well.
Do all frameworks run locally?
Many open-source frameworks like LangGraph, CrewAI, AutoGen, Dify, Langflow, OpenHands, and Camel AI can run locally with Ollama. OpenAI Agents SDK, Agentic AI, and some enterprise options typically rely on cloud APIs.
What do you need to get started?
Python basics, a local model via Ollama or LM Studio, patience, and depending on the framework, Docker.
When is OpenHands worth using?
OpenHands is ideal if you specifically want to change code in a repository, run tests, or prepare pull requests.
What’s the difference between LangGraph and CrewAI?
LangGraph models agents as graphs with states and decision edges. CrewAI works with roles and tasks that agents take on. LangGraph is more flexible, CrewAI is easier to get started with.
What’s the difference between Dify and Langflow?
Both are visual platforms. Dify is a comprehensive platform for chatbots, RAG, and agents with its own infrastructure. Langflow is a drag-and-drop interface for LangChain components. Dify is more all-in-one, Langflow is more modular.
Do you need a GPU for AI agents?
Not necessarily. If you use cloud APIs, a regular computer is enough. For local models via Ollama, a GPU significantly speeds up inference. Since agents often consume many tokens, a GPU is recommended.
Which model do you need for AI agents?
For tool calling and multi-step workflows, you need a capable model. Llama 3.1 8B is a good starting point. For more complex agents, 13B or larger models are recommended. In the cloud, GPT-4o or Claude are suitable.
Can you use multiple frameworks at once?
Yes. You could use CrewAI for role-based workflows and LangGraph for state-based workflows in parallel. Just make sure both can access the same Ollama instance.
What is multi-agent?
Multi-agent means multiple agents with different roles working together. CrewAI and AutoGen are examples. One agent writes code, another reviews it, a third documents it.
What does running a framework cost?
Open-source frameworks are free. Costs come from hardware, electricity, and optionally cloud APIs. With Ollama and local models, running costs are just electricity.
Can you run frameworks in Docker?
Yes. Dify, Langflow, OpenHands, and many other frameworks offer official Docker images. This is especially useful when you want to run multiple services isolated.
Where do I find the fundamentals on AI agents?
In the AI Agent Fundamentals overview, you’ll find articles on What is an AI Agent, Chatbot vs. AI Agent, and Tool Calling.