Agent frameworks and SDKs compared
Sixteen ways to build or run an agent: 14 code frameworks and SDKs, one low-code platform and one ready-made assistant. Every feature mark was checked against official docs on 1 October 2026, and a question mark means we could not confirm it.
Feature matrix
| Framework | Languages | Any model | MCP client | MCP server | Multi-agent | Durable runs | Human approval | Tracing | Evals | GitHub stars |
|---|---|---|---|---|---|---|---|---|---|---|
| OpenAI Agents SDK | Python | Yes | Yes | Not confirmed | Yes | Yes | Yes | Yes | No | 29.8k |
| OpenAI Agents SDK (JS) | TypeScript, JavaScript | Yes | Yes | Not confirmed | Yes | Not confirmed | Yes | Yes | No | 3.9k |
| Claude Agent SDK | Python, TypeScript | No | Yes | Not confirmed | Yes | Not confirmed | Yes | Yes | No | 8.2k |
| Google ADK | Python, TypeScript, Go, Java, Kotlin | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | 21.7k |
| Microsoft Agent Framework | Python, C#, .NET, Go | Yes | Yes | Not confirmed | Yes | Yes | Yes | Yes | Not confirmed | 13.9k |
| LangGraph | Python, TypeScript, JavaScript | Yes | Yes | Not confirmed | Yes | Yes | Yes | Yes | No | 42.6k |
| CrewAI | Python | Yes | Yes | Not confirmed | Yes | Yes | Yes | Yes | Not confirmed | 59.3k |
| LlamaIndex | Python, TypeScript | Yes | Yes | Yes | Yes | Yes | Yes | Not confirmed | Not confirmed | 52.4k |
| Pydantic AI | Python | Yes | Yes | Not confirmed | Yes | Yes | Yes | Yes | Yes | 20.3k |
| Mastra | TypeScript | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | 28.5k |
| Vercel AI SDK | TypeScript | Yes | Yes | Not confirmed | Yes | Yes | Yes | Not confirmed | Not confirmed | 27.1k |
| smolagents | Python | Yes | Yes | Not confirmed | Yes | Not confirmed | Not confirmed | Not confirmed | No | 29.6k |
| Agno | Python | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Not confirmed | 42.4k |
| Strands Agents | Python, TypeScript | Yes | Yes | Not confirmed | Yes | Not confirmed | Yes | Yes | Yes | 8.6k |
| Dify | Yes | Yes | Yes | Yes | Yes | Yes | Not confirmed | Not confirmed | 157.7k | |
| OpenClaw | TypeScript | Yes | Not confirmed | Not confirmed | Yes | Yes | Not confirmed | Not confirmed | No | 391.1k |
Which one to pick
Python, long runs that must survive restarts → LangGraph or Pydantic AI
LangGraph checkpoints a graph to Postgres or SQLite and resumes after failures. Pydantic AI plugs into seven durable engines, including Temporal and DBOS, and validates every output against a type.
TypeScript, one stack for everything → Mastra
Mastra ships agents, suspend-and-resume workflows, memory, scorers and an MCP server in one package. If you mainly need tool loops behind a streaming Next.js UI, the Vercel AI SDK is the narrower choice.
Azure, .NET or leaving AutoGen → Microsoft Agent Framework
It is the official successor to AutoGen and Semantic Kernel, with Python and C# SDKs, checkpointed graph workflows and OpenTelemetry built in.
Several languages or Google Cloud → Google ADK
ADK is the only option here with Java and Kotlin SDKs next to Python, TypeScript and Go. It also has all eight features confirmed, evals included.
Claude working on real files → Claude Agent SDK
It embeds the Claude Code harness, with file and shell tools, permissions, hooks and subagents, in your own program. It runs Claude models only, through Anthropic, Bedrock, Google Cloud or Foundry.
No framework code → Dify, or OpenClaw for a ready assistant
Dify builds agents and workflows in a visual editor and publishes them as APIs or MCP servers. OpenClaw is a finished assistant you host and message from chat apps, not something you build on.
Every framework in the table
OpenAI Agents SDK
OpenAI's open-source Python SDK for agents, kept deliberately thin, with handoffs, guardrails, sessions and tracing built in. It defaults to OpenAI models but can call other providers, and gets durable runs through a Temporal integration that went GA in March 2026.
OpenAI Agents SDK (JS)
The TypeScript and JavaScript port of OpenAI's Agents SDK, with the same agents, handoffs, approvals and tracing. It adds realtime voice agents that run in the browser over WebRTC, while the Temporal durability integration remains Python-only.
Claude Agent SDK
Anthropic's SDK puts the harness behind Claude Code, with file and shell tools, permissions, hooks, subagents and skills, inside your own Python or TypeScript program. Unlike the rest of this list it runs Claude models only, and Anthropic's Commercial Terms govern its use.
Google ADK
Google's Agent Development Kit, with SDKs for Python, TypeScript, Go, Java and Kotlin and evaluation built into the core. It is Gemini-first but reaches other models through adapters, and deploys to Agent Runtime on Google Cloud.
Microsoft Agent Framework
Microsoft's agent framework for Python and .NET, built by the Semantic Kernel and AutoGen teams and generally available since April 2026. It pairs typed agents with graph workflows that checkpoint, resume and wait for people, and a Go SDK is in public preview.
LangGraph
LangChain's low-level library for agents built as explicit graphs, with checkpointed state and interrupts at its core. LangChain v1's create_agent runs on top of it for when a prebuilt loop is enough.
CrewAI
A Python framework for crews of role-playing agents that delegate tasks to each other, orchestrated by event-driven Flows. Flows add persistence and human feedback, and the CLI deploys straight to the managed CrewAI AMP service.
LlamaIndex
LlamaIndex pairs its document parsing and retrieval tools with event-driven Agent Workflows in Python and TypeScript. Workflows can pause for human input, persist state, and be served as MCP servers or deployed to LlamaCloud.
Pydantic AI
A Python agent framework from the Pydantic team, built around type hints, validated structured outputs and dependency injection. It stands out for durable execution on seven engines, from Temporal to Airflow, and ships Pydantic Evals alongside.
Mastra
A TypeScript framework that bundles agents, workflows with suspend and resume, memory, evals and the Studio UI in one package. It drops into Next.js, Express or Hono apps; the core is Apache-2.0, while ee/ directories use an enterprise licence.
Vercel AI SDK
Vercel's provider-agnostic TypeScript toolkit, where ToolLoopAgent is the main agent class as of AI SDK 7. It ties tool loops to streaming UI helpers, and WorkflowAgent adds runs that survive restarts.
smolagents
Hugging Face's small Python library, about a thousand lines, whose CodeAgent writes its actions as Python code. That code runs in sandboxes such as Modal, E2B or Docker, and agents and tools can be shared on the Hub.
Agno
A Python framework and runtime: the SDK builds agents, teams and workflows, and AgentOS serves them as a FastAPI app on your own infrastructure. Sessions, memory and traces stay in your database, and a hosted Control Plane UI connects from the browser.
Strands Agents
An AWS SDK for Python and TypeScript where the model drives the loop and the agent runs in your own process, with no control plane. Amazon Bedrock is the default provider, and Bedrock AgentCore Runtime is the deployment path.
Dify
A self-hostable platform for building LLM agents and workflows in a visual editor instead of framework code. Since v1.6.0 it speaks MCP in both directions, so an app can call MCP tools and be published as an MCP server.
OpenClaw
A ready-made, self-hosted AI assistant you message from WhatsApp, Telegram, Slack, Discord, iMessage and more. It is not a framework for building your own agent, and by design there is no hosted service in the middle.
Questions about agent frameworks
What is the best AI agent framework in 2026?
There is no single winner: the right pick follows your language, your cloud and how long your runs last. Google ADK and Mastra are the only two in our table with all eight features confirmed. Pydantic AI, LangGraph and CrewAI lack confirmation on only one or two, so filter by language first and compare from there.
Which agent frameworks support MCP?
All 14 code frameworks here, and Dify, can call tools on MCP servers. Fewer document exposing your own agent as an MCP server: Google ADK (to_mcp_server()), LlamaIndex (workflow_as_mcp), Mastra (MCPServer), Agno (through AgentOS) and Dify. We could not confirm MCP support for OpenClaw in its docs.
Is AutoGen deprecated?
AutoGen is in maintenance mode, and its last release was python-v0.7.5 in September 2025. Microsoft Agent Framework 1.0, generally available since 3 April 2026, is the official successor to both AutoGen and Semantic Kernel. Semantic Kernel v1 keeps getting critical fixes for at least a year after that date.
Is OpenAI Agent Builder shutting down?
Yes. On 3 June 2026 OpenAI announced it is winding down Agent Builder and its Evals product, and both close on 30 November 2026. ChatKit stays, but you run its backend yourself, for example with the Agents SDK. The open-source Agents SDK is a separate library and still ships releases; 0.22.3 came out on 17 September 2026.
Which agent frameworks support durable execution?
LangGraph (checkpointers), Pydantic AI (seven engines from Temporal to Airflow), Microsoft Agent Framework (Durable Extension), the Python OpenAI Agents SDK (Temporal) and the Vercel AI SDK (WorkflowAgent) all document it. Mastra and Dify can pause a workflow and resume it later. For the Claude Agent SDK, Strands Agents and the OpenAI JS SDK we found resumable sessions but no durable engine.
What does a question mark in the table mean?
It means we could not confirm the feature in official documentation on 1 October 2026, not that it is missing. An empty circle means the library has no built-in support, even when a separate product covers it, as LangSmith does for LangGraph evals.