There are a dozen serious AI agent frameworks now, and the differences are real — chains vs graphs vs role-based crews vs SDKs. Here is a neutral index by language, design paradigm, license, and what each is genuinely best at. These are open-source libraries, so there are no prices — just a decision matrix.
The matrix
| Framework | Languages | Paradigm | License | Best for |
|---|---|---|---|---|
| LangChain | Python · JS/TS | Chains / pipelines | Open-source | General-purpose LLM apps; the broadest integration ecosystem |
| LangGraph | Python · JS/TS | Graph / stateful | Open-source | Controllable, stateful, multi-step and multi-agent flows with human-in-the-loop |
| LlamaIndex | Python · TS | Data / RAG-centric | Open-source | Agents that reason over your own data (retrieval-augmented generation) |
| CrewAI | Python | Role-based crews | Open-source | Spinning up a team of role-playing agents quickly |
| Microsoft AutoGen (AG2) | Python | Conversational multi-agent | Open-source | Multi-agent conversation and research-style collaboration patterns |
| Microsoft Semantic Kernel | C# · Python · Java | SDK / plugins | Open-source | Embedding AI into enterprise and .NET applications |
| OpenAI Agents SDK | Python · JS/TS | Chains / pipelines | Open-source | Lightweight production agents with tools and handoffs (successor to Swarm) |
| Pydantic AI | Python | Type-safe | Open-source | Pythonic, type-checked agents with structured outputs |
| Haystack | Python | Data / RAG-centric | Open-source | Production search and RAG pipelines, with agent support (deepset) |
| Google ADK | Python · Java | SDK / plugins | Open-source | Building and deploying agents on Gemini / Vertex AI and Google Cloud |
| Vercel AI SDK | TypeScript | SDK / plugins | Open-source | AI features and agents in web apps, with streaming UI |
| n8n | No-code (+ JS) | No-code visual | Source-available · self-host | No-code agents (native AI / LangChain nodes) wired into real automations |
Quick picks
- You want the broadest ecosystem and maximum flexibility → LangChain
- You need controllable, stateful, multi-step flows → LangGraph
- The agent's value is RAG over your own data → LlamaIndex or Haystack
- You want a team of role-playing agents fast → CrewAI or AutoGen
- You're a .NET / enterprise shop → Semantic Kernel
- You want type-safe, Pythonic agents → Pydantic AI
- You're shipping agents inside a web / TypeScript app → Vercel AI SDK
- You want no-code agents wired into automations → n8n
📚 More from The 2026 AI Stack Index: Automation Tools · Agent Frameworks · Vector Databases · LLM Observability · LLM Gateways
This is a neutral, no-affiliate reference — no prices (they go stale), no rankings-for-pay. The full, always-updated interactive version with FAQs and the rest of the AI-stack indexes lives at aiprosol.com/agent-frameworks. Disclosure: I run Aiprosol, an automation consultancy — the index doesn't favour anyone.
Top comments (1)
The useful comparison point for agent frameworks is not only orchestration style. I would also compare how each one handles skills, tool permissions, memory boundaries, human approvals, retries, and evidence logs. Those are the parts that decide whether a demo turns into a reliable operating workflow.