Which AI agent framework should you use? The pick for each job (Nov 2026)

Which AI agent framework should you use? Pick by the job, not by popularity. For a durable, auditable production agent with human approval steps, use LangGraph. For the fastest multi-agent prototype, use CrewAI. If you build around one model vendor, use that vendor's kit: the OpenAI Agents SDK, the Claude Agent SDK or Google ADK. If you were on AutoGen, move: Microsoft archived it and points new users to Microsoft Agent Framework. If you'd rather not code, use n8n. Every version and fact below was read on the project's own docs and GitHub pages on November 3, 2026.

Which framework fits my job?

Find the row that matches what you're building. The versions are each project's latest GitHub release as of November 3, 2026, so you can see which ones are actively shipping.

• You need it in production, with an audit trail and a human approving risky steps: LangGraph. Its own docs list durable execution, streaming and human-in-the-loop as the point of the framework.
• You want a multi-agent demo working this afternoon: CrewAI. You describe agents by role and give them tasks; Flows add the structured control for when it goes live.
• You're all-in on GPT models: the OpenAI Agents SDK, which OpenAI describes as having "very few abstractions".
• You want an agent that reads files, runs commands and edits code on Claude: the Claude Agent SDK, which gives you "the same tools, agent loop, and context management that power Claude Code".
• You need Go, Java or Kotlin, or you're on Google Cloud: Google ADK.
• You're on AutoGen, Semantic Kernel, Azure or .NET: Microsoft Agent Framework.
• The logic is simple but it touches a dozen SaaS apps, and you don't want to code: n8n.

What happened to AutoGen?

Microsoft archived it. The README on Microsoft's AutoGen GitHub page now opens with a caution box: "Project archived and no longer maintained". It says Microsoft "is not providing bug fixes, security fixes, new features, or support for this project" and tells new users to start with Microsoft Agent Framework. The last AutoGen release on GitHub is python-v0.7.5, published September 30, 2025.

Microsoft's docs call Agent Framework "the direct successor, created by the same teams" behind Semantic Kernel and AutoGen. It works in .NET, Python and Go (Go is a public preview), and Microsoft lists Foundry, Anthropic, Azure OpenAI, OpenAI and Ollama among its model providers. Microsoft publishes a migration guide from AutoGen.

AG2 is a different project. It's the community continuation of the original AutoGen code, it's still releasing (v1.2.2 on October 27, 2026, Apache-2.0), and it's what you want if you like the classic conversational style. When a tutorial says "AutoGen", check which one it means.

Is LangGraph still the safe pick for production?

Yes, if you're willing to design the flow yourself. LangChain's docs describe LangGraph as "a low-level orchestration framework and runtime for long-running, stateful agents". You define the work as a graph of shared state, nodes and edges, and you can mix fixed steps with LLM-driven ones in the same graph.

• What you get: durable execution (a run can pick up after a crash or a pause), human-in-the-loop (pause, inspect, change state, resume), memory across sessions and streaming.
• What it costs you: the same page says LangGraph "does not prescribe prompts or agent architecture". That's freedom, and it's also work.
• You don't need LangChain: the docs say "LangGraph does not require LangChain". If you want a ready-made agent loop first, LangChain points you to its Deep Agents or LangChain agents and says to use LangGraph "when you need explicit control over state and control flow".

Is CrewAI good enough for production now?

For a lot of jobs, yes. The quick "crew" of role-playing agents is still the fastest start, but CrewAI's own docs now tell you to build production apps around Flows: structured, event-driven workflows that "manage state and control execution", with a Crew handling a step that needs agents working together.

A common path is to prototype in CrewAI and move the parts that need tight control to LangGraph once the requirements settle. Our LangGraph vs CrewAI comparison goes through that choice in detail.

Can I use the OpenAI Agents SDK with Claude, Gemini or a local model?

You can, but it's built for OpenAI first. The SDK's model docs say it "currently includes Any-LLM and LiteLLM as best-effort, beta adapter integrations", and that some tool features "are supported only with OpenAI Responses models". They also suggest turning tracing off if you don't have an OpenAI platform key.

So if you plan to switch between model vendors, a model-neutral framework (LangGraph, CrewAI, Pydantic AI, Microsoft Agent Framework or Google ADK) will fight you less. If you're on GPT models anyway, the OpenAI SDK is the shortest path. Its core pieces are agents with instructions and tools, handoffs that delegate to other agents, and guardrails that check inputs and outputs.

What about the Claude Agent SDK, Google ADK and Pydantic AI?

• Claude Agent SDK (Anthropic): "Build production AI agents with Claude Code as a library." It comes with built-in tools to read, write and edit files, run commands and search the web, plus hooks, subagents, MCP, permissions and sessions, in Python and TypeScript. It's built around Claude, and its use falls under Anthropic's Commercial Terms of Service. Pick it for coding and file-heavy agents on Claude.
• Google ADK: "the open-source agent development framework that lets you build, debug, and deploy reliable AI agents", in Python, TypeScript, Go, Java and Kotlin. Google says it "can work with almost any generative AI model", though its examples use Gemini.
• Pydantic AI: "the Python AI SDK: a typed, extensible agent loop with every model a string swap away", with a fallback model for when a provider fails. A good fit if your team already writes typed Python.

Do I need to code to build an agent?

No. n8n's AI Agent node lets you "connect a chat model and one or more tools, and the agent decides which tools to call". You must connect at least one tool. n8n's integrations directory lists 2,000+ integrations, which is the real reason to use it: the hard part of most business agents is the glue, not the reasoning.

Two things to know before you commit. n8n's code is under its Sustainable Use License, a source-available license with limits, not a standard open-source license, so read it if you plan to resell n8n as a service. And there's a self-hosted Community Edition on GitHub alongside the paid cloud plans on n8n's pricing page. More in our n8n for AI agents guide.

Can I run my agent on an open model I control?

Yes. The model-neutral frameworks don't tie you to one vendor: Google's ADK docs mention adapters for locally run models, Microsoft lists Ollama among Agent Framework's model providers, and the OpenAI SDK can reach other providers through its beta adapters. The question becomes hardware: whether the model fits in your GPU's memory. Check a model against your card with the VRAM calculator, and see how to run a local AI agent for the setup.

If your own card is too small, renting one by the hour is the cheap way to test before you buy. Today's per-GPU prices from Vast and Runpod are on /cloud-gpu.

New to agents? Start with what AI agents are and can't do. Going deeper: LangGraph vs CrewAI and n8n for AI agents. Running them yourself: how to run a local AI agent.