Level : Advanced
AI Agent Frameworks Compared: LangGraph, CrewAI, ADK, Agents SDK and More
By The Info Agent IA editorial team ·
An AI agent framework provides the plumbing that connects a language model to tools, memory and other agents; as of October 22, 2026, the strongest active options are LangGraph, CrewAI, Microsoft Agent Framework, Google ADK, the OpenAI Agents SDK, Pydantic AI and Mastra, each suited to a different profile. Before comparing them, rule out the projects that have been shut down or replaced: AutoGen is archived, Semantic Kernel has become Microsoft Agent Framework, and several platforms are closing or changing names.
What a framework gives you, and when to skip one
At heart, an agent is a loop: the model reads the goal, picks a tool, gets the result and decides on the next step. A few dozen lines of code are enough to run that loop against a provider’s API. A framework comes in later, when the loop has to become dependable:
- State and recovery: saving where the agent is, resuming after a crash, replaying a step.
- Memory: keeping the context of a conversation or a user from one session to the next.
- Orchestration: chaining several steps or several agents, with branches and loops.
- Integrations: connectors to model providers, support for MCP for tools and, increasingly, for the A2A protocol so agents can talk to each other.
- Monitoring: traces of every call, costs and evaluations, known as LLM observability.
If your need boils down to one task and a few tools, a visual platform may be enough; our tutorial on building an AI agent without code covers that route. For the big picture, from no-code to code, see our guide to building an AI agent.
Start with status: what has been retired or replaced
The landscape shifted a lot in 2025 and 2026. Several names still common in tutorials are no longer sensible choices for a new project.
The practical upshot: a 2024 tutorial built on AutoGen or Semantic Kernel will have you write code you’ll later need to migrate. Always check the date of an example and the project’s status on its profile.
Active code frameworks compared
Here are the ten code frameworks most often shortlisted, with their characteristics as listed on their profiles (type, license, MCP support, multi-agent support, required skill level):
| Tool | Type | License | MCP | Multi-agent | Level |
|---|---|---|---|---|---|
| LangGraph | Code | MIT | Yes | Yes | Advanced |
| CrewAI | Code | MIT | Yes | Yes | Intermediate |
| Microsoft Agent Framework | Code | MIT | Yes | Yes | Intermediate |
| Google Agent Development Kit (ADK) | Code | Apache-2.0 | Yes | Yes | Intermediate |
| OpenAI Agents SDK | Code | MIT | Yes | Yes | Intermediate |
| Claude Agent SDK | Code | proprietary | Yes | Yes | Intermediate |
| Pydantic AI | Code | MIT | Yes | Yes | Intermediate |
| Mastra | Code | Apache-2.0 | Yes | Yes | Intermediate |
| LlamaIndex | Code | MIT | Yes | Yes | Intermediate |
| smolagents | Code | Apache-2.0 | Yes | Yes | Beginner |
All of them support MCP and multi-agent architectures. The differences lie elsewhere: language, degree of control, ties to a model provider and the paid ecosystem around the open source core.
Profiles, family by family
Graph-based frameworks: LangGraph and LangChain
LangGraph models an agent as a graph: nodes (call the model, run a tool, ask for approval) linked by transitions, with shared state saved at every step. It’s the most explicit framework on the list: you see and control every possible path, and you can pause the agent for human approval, then resume it. The trade-off is a steeper learning curve, hence the “advanced” level on its profile. MIT license, Python and TypeScript versions.
LangChain, from the same company, now serves as a higher-level layer: its agents are built on LangGraph, and it brings a huge library of integrations. The related paid offering is LangSmith (observability, evaluation, deployment): a free Developer tier and Plus at $39 per seat per month as of October 22, 2026, with deployment billed by usage.
Who it’s for: teams building long-running or critical agents that need error recovery, human approval and an auditable record of every transition.
Agent teams: CrewAI
CrewAI starts from a metaphor: a crew of agents, each with a role, a goal and tools, splitting up tasks. You describe who does what, often in configuration files, and the framework organizes the collaboration; it also offers “flows” for more deterministic sequences. It’s a standalone Python framework, independent of LangChain, compatible with Python 3.10 to 3.13, under MIT. The commercial CrewAI AMP offering adds managed deployment and governance; its free Basic plan is capped at 50 workflow executions a month.
Who it’s for: people who think about their problem in roles (researcher, writer, reviewer) and want a working multi-agent system fast, without drawing every transition.
The big vendors’ SDKs
Microsoft Agent Framework is the direct successor to AutoGen and Semantic Kernel, built by the same teams. Version 1.0 for Python and .NET shipped on April 2, 2026, a Go SDK is in public preview, and the license is MIT. It combines AutoGen’s conversational agents with Semantic Kernel’s enterprise building blocks, with optional paid hosting on Microsoft’s cloud. It’s the obvious choice for a .NET team or an Azure shop, and the migration target for both older projects.
Google Agent Development Kit (ADK) comes in Python, TypeScript, Go and Java, in separate Apache-2.0 repositories. It supports the A2A protocol and can be deployed, as a paid option, on Google Cloud (Agent Runtime, Cloud Run, GKE). It works with models other than Gemini, but its integration runs deepest within Google’s ecosystem.
OpenAI Agents SDK, the successor to the experimental Swarm project, bets on simplicity: a handful of primitives (agents, tools, handoffs between agents, guardrails) in Python and TypeScript, under MIT. The SDK is free; only model calls are billed. It’s also the migration path OpenAI recommends for workflows built in Agent Builder, which closes on November 30, 2026.
Claude Agent SDK, formerly the Claude Code SDK and renamed in September 2025, exposes the agent loop that powers Claude Code: reading files, running commands, searching, sub-agents. It’s free but governed by Anthropic’s commercial terms, and it works only with Claude models. A great fit for agents that handle code or files; not the one to pick if you want to switch providers.
On the AWS side, Strands Agents (Apache-2.0, Python and TypeScript) plays a similar role, with natural deployment on Bedrock AgentCore.
Who they’re for: anyone who has already chosen a cloud or model provider: the in-house SDK removes friction, at the cost of tighter lock-in.
Lightweight, typed Python frameworks
Pydantic AI, from the team behind the Pydantic validation library, brings the same philosophy to agents: typed, validated inputs and outputs, with guaranteed structured output. Version 1.0 shipped in September 2025, under MIT. The paid offering doesn’t host agents: it’s Logfire, an observability service (Team at $49 a month, Growth at $249 a month as of October 22, 2026), and the OpenTelemetry instrumentation works with other tools too.
smolagents, from Hugging Face (Apache-2.0), is the most approachable option on the list, rated “beginner” on its profile. Its signature feature, CodeAgent, writes its actions as Python code rather than as JSON tool calls; execution can be sandboxed in Docker or with third-party services.
Agno, formerly Phidata, offers a free SDK and runtime (AgentOS), with a Pro plan at $150 a month to connect its control plane to a deployment.
Who they’re for: Python developers who want to write ordinary, testable application code without adopting a whole architecture.
The TypeScript world: Mastra and the Vercel AI SDK
Mastra is a full TypeScript framework: agents, workflows, memory, evaluations. Version 1.0 shipped in January 2026. Watch the dual license: the core is Apache-2.0, but code in the ee/ folders falls under an enterprise license for production use.
The Vercel AI SDK (Apache-2.0, npm package “ai”) is more a library than a framework: one interface to many providers, widely used in web apps. By default, calls go through Vercel AI Gateway, a separate billed service, but you can call providers directly.
Who they’re for: web and Node.js teams that want a single tech stack from front end to back end.
Documents and RAG: LlamaIndex and Haystack
LlamaIndex (MIT) made its name indexing documents and remains geared to agents that work on data. The company now focuses on LlamaParse, its paid document-parsing platform, and the TypeScript version of the framework is archived. Haystack, from deepset (Apache-2.0), released version 3 in July 2026 and targets production document-search pipelines, with an enterprise offering.
Who they’re for: projects where most of the work is finding the right information in a body of documents before reasoning about it.
Special cases: DSPy and cloud platforms
DSPy (MIT), born at Stanford, works differently from the rest: instead of writing prompts, you describe modules and metrics, and the framework optimizes the prompts automatically. Powerful if you have evaluation data, puzzling if you don’t; version 3.0 shipped in August 2025.
Amazon Bedrock AgentCore and Gemini Enterprise Agent Platform aren’t frameworks in the strict sense: they host, secure and monitor agents written with whatever framework you like. They’re paid and aimed at advanced users; consider them once you’ve picked your framework.
The verdict: which framework for whom?
| Your situation | First option to evaluate | Why |
|---|---|---|
| Long-running, critical agent with human approvals | LangGraph | Saved state, recovery, control over every transition |
| Role-based agent team, quick prototype | CrewAI | Declarative roles and tasks, native multi-agent |
| .NET or Azure shop, legacy AutoGen or Semantic Kernel code | Microsoft Agent Framework | Official successor, Python and .NET |
| Project on Google Cloud, need for A2A | Google ADK | Four languages, managed Google deployment |
| Simple agent on OpenAI models, migrating from Agent Builder | OpenAI Agents SDK | Few concepts, official migration target |
| Agent that handles code and files, Claude models | Claude Agent SDK | Claude Code’s loop, built-in file tools |
| Typed, testable Python application | Pydantic AI | Validated outputs, ordinary Python code |
| End-to-end TypeScript stack | Mastra or Vercel AI SDK | Full framework or lightweight library |
| Agents over a large document collection | LlamaIndex or Haystack | Indexing and search at the core |
| First experiment, learning | smolagents | Few concepts, beginner level |
Check yourself Your team maintains an internal assistant written with AutoGen in 2024. It works. Should you migrate it?
Yes, and plan for it. AutoGen is archived: no more fixes of any kind, including security fixes, and Microsoft advises against production use. The logical target is Microsoft Agent Framework, built by the same team and reusing its concepts. There’s no need to break what works overnight, but any vulnerability found in AutoGen or its dependencies will stay open.
The criteria that really decide it
Language. Often the first filter. In Python, almost everything is available. In TypeScript: LangGraph, Mastra, the Vercel AI SDK, the OpenAI Agents SDK, Google ADK and Strands. In .NET: Microsoft Agent Framework. In Go and Java: Google ADK, plus Go in preview at Microsoft.
Ties to a model provider. Most frameworks accept several providers. The Claude Agent SDK works only with Claude; the OpenAI and Google SDKs accept other models, but their most advanced features are designed around their own ecosystems. If the freedom to switch models matters to you, choose a neutral framework.
Protocols. Every framework in the table supports MCP, which lets you plug in standardized tools; if you plan to expose your own tools, our guide to building an MCP server shows how. A2A, for agent-to-agent communication, is championed by Google ADK and BeeAI.
License. MIT and Apache-2.0 give you wide latitude, including for commercial use. Read the exceptions: Mastra’s enterprise code, the Claude Agent SDK’s commercial terms.
Real cost. The framework is almost always free; what costs money is model calls, then hosting and observability. An agent that loops ten times on a task burns ten times the tokens of a single call.
Ability to evaluate. A framework that makes tracing and testing easy will save you more time than any flashy feature. Our guide to evaluating an AI agent walks through the method.
To filter and compare every profile, including those not covered here (Atomic Agents, BeeAI, CAMEL, Langroid, OpenHands and others), use our agent framework comparison tool.
Frequently asked questions
What is the best framework for building an AI agent?
There isn't a single one: it depends on your language, your model provider and how much control you want. LangGraph suits complex flows with error recovery, CrewAI suits role-based agent teams, Pydantic AI suits typed Python applications and Mastra suits TypeScript projects. The OpenAI, Google and Microsoft SDKs are the natural picks if you already work with those vendors.
Should I still use AutoGen or Semantic Kernel?
Not for a new project. AutoGen is archived and no longer receives fixes, including security fixes; Semantic Kernel has become Microsoft Agent Framework, with v1.x maintained for at least a year after Agent Framework's April 2026 release. Microsoft points users of both toward Agent Framework.
Are AI agent frameworks free?
Most are open source and free, under the MIT or Apache-2.0 license. You pay the model provider for API calls, and possibly for a hosting or observability service such as LangSmith (Plus at $39 per seat per month as of October 22, 2026) or Pydantic Logfire. The Claude Agent SDK is free but governed by Anthropic's commercial terms.
Do I actually need an agent framework?
No. For a simple agent, a few calls to a model API with tool calling are enough, and a no-code platform may do the job. A framework pays off when you need to manage state, memory, multiple agents, recovery after errors or production monitoring.