The latest AI agent frameworks in 2026, and which one to build on

The AI agent frameworks most teams choose from in late 2026 are LangGraph, CrewAI, Microsoft Agent Framework, Google’s ADK, the OpenAI Agents SDK, the Claude Agent SDK, AWS Strands Agents, Pydantic AI and Mastra. Every one of them now supports MCP, the open standard for connecting an agent to tools, so the protocol matters more than the framework.

Below are their latest versions, what changed this year, and how we pick one when we build an agent for a client. We checked every version on PyPI, npm or the maker’s GitHub releases on 6 November 2026.

Which AI agent frameworks are the latest in 2026?

Framework	Maker	Languages	Latest version
LangGraph	LangChain	Python, TypeScript	1.2.20, 5 Nov 2026
CrewAI	CrewAI	Python	1.15.35, 3 Nov 2026
Microsoft Agent Framework	Microsoft	Python, .NET	1.24.0, 30 Oct 2026
Agent Development Kit (ADK)	Google	Python, TypeScript, Go, Java, Kotlin	2.13.0, 23 Oct 2026
OpenAI Agents SDK	OpenAI	Python, TypeScript	0.24.0, 15 Oct 2026
Claude Agent SDK	Anthropic	Python, TypeScript	0.2.178, 4 Nov 2026
Strands Agents	AWS	Python, TypeScript	1.64.0, 4 Nov 2026
Pydantic AI	Pydantic	Python	2.66.0, 4 Nov 2026
Mastra	Mastra	TypeScript	1.82.0, 4 Nov 2026
LlamaIndex	LlamaIndex	Python, TypeScript	0.14.25, 21 Sep 2026
Dify	LangGenius	Visual builder	1.19.0, 4 Nov 2026

Versions are as of 6 November 2026. Where a framework ships in several languages, the version is the Python package, and for Mastra it is @mastra/core.

Most of these ship a new version every week or two, so treat the numbers as a snapshot. The OpenAI Agents SDK and the Claude Agent SDK are still below 1.0.

All of them are open source, with two catches. The TypeScript Claude Agent SDK is under Anthropic’s commercial terms, and Dify uses its own licence based on Apache 2.0.

Which frameworks support MCP and A2A?

Framework	MCP	A2A
LangGraph	Yes	Partly, through LangGraph’s Agent Server
CrewAI	Yes	Yes
Microsoft Agent Framework	Yes	Pre-release packages only
Google ADK	Yes	Yes
OpenAI Agents SDK	Yes	Not in the docs
Claude Agent SDK	Yes	Not in the docs
Strands Agents	Yes	Yes
Pydantic AI	Yes	In a separate package, FastA2A, since version 2
Mastra	Yes	Yes
LlamaIndex	Yes	Not confirmed
Dify	Yes	Not built in

“Not in the docs” means we found no A2A support in the maker’s documentation. Neither maker says it is unsupported.

What are MCP and A2A?

MCP, the Model Context Protocol, connects an agent to tools and data: your CRM, your inbox, a database. Anthropic released it, and it moved to the Agentic AI Foundation, part of the Linux Foundation, on 9 December 2025.

A2A, the Agent2Agent protocol, lets one agent hand work to another, even when the two are built on different frameworks. Google started it, version 1.0 shipped on 12 March 2026, and it joined the same foundation on 27 August 2026.

MCP is the one a business needs now. A2A matters once you run several agents built by different teams or vendors that must pass work between them.

If you sell software, MCP is also how ChatGPT and Claude use your product for a buyer. We wrote up why SaaS products need an MCP server.

What changed in AI agent frameworks in 2026?

• MCP went stateless. The 2026-07-28 spec, released on 28 July, dropped the session handshake. An MCP server can now run like any other web API, which makes it simpler to host and scale.
• A2A reached 1.0. Version 1.0 shipped on 12 March and 1.0.1 on 28 May. AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow sit on its technical steering committee.
• Microsoft merged its two frameworks. Microsoft Agent Framework 1.0 shipped on 2 April as the successor to AutoGen and Semantic Kernel. AutoGen is in maintenance mode and being archived.
• Google shipped ADK 2.0. The Python release on 19 May added an engine for multi-agent workflows. TypeScript followed on 21 August.
• Pydantic AI 2.0 moved A2A out. Version 2.0 shipped on 23 June. A2A now lives in a separate package, so older guides that say it is built in are out of date.
• TypeScript caught up. Mastra reached 1.0 on 20 January, and the TypeScript version of Strands Agents on 30 April.
• Agents got their own workspaces. The OpenAI Agents SDK added Sandbox Agents in beta on 15 April, which give an agent a persistent, isolated workspace to work in.

What is OpenClaw?

OpenClaw is a self-hosted personal AI assistant. It connects chat apps such as WhatsApp, Telegram, Slack and Discord to an agent running on your own machine.

Peter Steinberger started it in November 2025, and it took the name OpenClaw on 29 January 2026, after being called Clawd and then Moltbot. It is MIT-licensed, works as both an MCP client and server, and its latest release is v2026.11.1, from 5 November 2026.

Steinberger joined OpenAI in February 2026. OpenClaw now sits with the OpenClaw Foundation, a non-profit announced on 8 July 2026, and the foundation says it is not an OpenAI product.

OpenClaw is a personal assistant you run for yourself. It is not a framework for building an agent into a company’s own tools.

Which other agent projects come up?

Project	What it is	Latest version
Hermes Agent	An open-source agent from Nous Research, MIT-licensed, with MCP client and server support	v0.22.7, 3 Nov 2026
Letta	Agents with long-term memory, now centred on Letta Code	Letta Code v0.34.41, 5 Nov 2026
NLWeb	Turns a website into something agents can query, and makes every site an MCP server	No tagged releases

What about autonomous coding agents?

Two of the frameworks above come from coding agents. The Claude Agent SDK is the agent loop and tools behind Claude Code, packaged as a library. The OpenAI Agents SDK’s Sandbox Agents give an agent its own workspace to edit files and run code.

We use Claude for coding in our own companies, with a person reviewing every merge. Claude co-wrote 70 of 98 site commits at Shhots AI in August 2026, and at Mockey a Claude routine reads open issues and pushes the fixes for a developer to review.

Which framework should a business build on?

Pick the model first, then the framework. For most business agents we use the model maker’s own SDK, because new model features usually land there first and there is less code between the agent and the model.

Your situation	What we would build on
One model, one job, such as answering leads	The model maker’s SDK: Claude Agent SDK, OpenAI Agents SDK or ADK
You want to switch models without a rewrite	LangGraph or Pydantic AI
A job runs for hours and must survive a restart	LangGraph, built for long-running state
Your team writes TypeScript	Mastra
You run on Azure and .NET	Microsoft Agent Framework
You run on AWS	Strands Agents
A non-developer must change the steps	Dify, a visual builder

Whatever we pick, these rules hold:

• The agent talks to your tools through MCP, so the tools outlast the framework.
• It runs in your own accounts, so you keep it if we stop.
• A person approves anything that sends, hires or moves money.

The framework is rarely why an agent fails. Most agents that stall do so because nobody owns them after launch. And many jobs never needed an agent at all, only a standing order: a known rule, run every time.

Where should you start?

Start with the job, not the framework. Pick one job your team repeats every week where the next step needs judgement, and write down the number it should move.

Bring that to the calendar below. We will tell you whether it needs an agent, which framework we would build it on, and how we would build and run it.