The AI Agent Frameworks Index
A neutral index of the AI agent frameworks developers actually reach for in 2026 — grouped by language, paradigm, and what each is genuinely best at. The agent ecosystem moves fast, so we describe each framework's design (chains, graphs, crews, SDKs) rather than version-specific feature claims, and we don't quote prices: these are mostly free, open-source libraries, and your real cost is the model API calls underneath. Pair this with the automation-tool index when you're deciding between writing code and wiring no-code.
Last reviewed June 13, 2026 · 12 frameworks · neutral & vendor-independent
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 |
Paradigms and languages, not version-specific features — the ecosystem moves fast and a stale feature claim is worse than none. Check each project's repo for the latest before committing.
How to read the paradigms
- Chains / pipelines
- Compose LLM calls, tools, and retrieval into linear or branching sequences. The most general starting point.
- Graph / stateful
- Model the agent as a state machine with cycles, memory, and human-in-the-loop checkpoints — more control over multi-step runs.
- Role-based crews
- Multiple specialised agents, each with a role and goal, collaborate to finish a task.
- Conversational multi-agent
- Agents (and tools) talk to each other to solve a problem — strong for research and emergent collaboration.
- SDK / plugins
- A library that embeds AI plus tool plugins and planners into your existing application code.
- Type-safe
- Structured, type-checked inputs and outputs — fewer silent failures when agents hit production.
- Data / RAG-centric
- Built around indexing and retrieving your own data so answers are grounded, not guessed.
- No-code visual
- Build agents by wiring nodes on a canvas — no programming required, and they drop straight into automations.
Which one should you pick?
- You want the broadest ecosystem and maximum flexibilityLangChain
- You need controllable, stateful, multi-step flowsLangGraph
- The agent's value is RAG over your own dataLlamaIndex or Haystack
- You want a team of role-playing agents fastCrewAI or AutoGen
- You're a .NET / enterprise shopSemantic Kernel
- You want type-safe, Pythonic agentsPydantic AI
- You're shipping agents inside a web / TypeScript appVercel AI SDK
- You want no-code agents wired into automationsn8n
Related head-to-heads
Make vs n8n
Make is a polished cloud-only visual builder with strong branching, billed per operation. n8n is open-source and self-hostable, th…
Read the comparison →Zapier vs n8n
Zapier is the most beginner-friendly option with the largest app-integration library, cloud-only and billed per task. n8n is open-…
Read the comparison →FAQs
What's the difference between an AI agent framework and a workflow automation tool?
An agent framework is a code library for building LLM-powered agents — reasoning loops, tool use, memory, and orchestration. A workflow automation tool (Zapier, Make, n8n) is a platform for wiring apps together, increasingly with AI steps bolted on. The lines blur — n8n appears on both lists — but frameworks give developers low-level control and live in your codebase, while automation tools give non-developers speed and a visual canvas. Many teams use both: a no-code tool for the plumbing and a framework for the hard reasoning.
Which AI agent framework is the best?
There's no single winner — it depends on the job. For the broadest ecosystem and flexibility, LangChain; for controllable, stateful flows, LangGraph; for retrieval over your own data, LlamaIndex or Haystack; for multi-agent collaboration, CrewAI or AutoGen; for enterprise and .NET, Semantic Kernel; for type-safe Python, Pydantic AI. Most production teams end up combining two — for example LangGraph for orchestration and LlamaIndex for retrieval.
Do I need a framework to build an AI agent?
No. You can build a capable agent by calling an LLM API directly in a loop that reads tool results and decides the next step. Frameworks add structure, integrations, memory, retries, and observability, which save real time as agents grow — but they also add a dependency and a learning curve. For a first prototype, a plain API loop is often the clearest way to understand what an agent actually is.
Are these frameworks free?
Most are free, open-source libraries. Your real costs are the underlying model API calls (billed per token) and the compute/hosting to run the agent — not the framework itself. n8n is source-available and self-hostable. As always, the framework is the cheap part; the model usage and the engineering to operate agents reliably are where the cost sits.
How current is this index?
The agent-framework ecosystem moves quickly, so we deliberately describe each framework's paradigm and supported languages rather than version-specific feature claims that would go stale. A last-reviewed date is shown at the top of the page; always check the project's own repository and docs for the latest before committing to one.