Best AI Agent Frameworks and Platforms (November 2026)

AI agent frameworks are software libraries for building applications that use models to select tools and complete tasks. Agent platforms provide services for running those applications. Shortlist the Vercel AI SDK for TypeScript apps, LangGraph for stateful workflows and Microsoft Agent Framework for .NET integration; choose a hosted runtime when your team wants to offload deployment and session management.

The key choices are your team's language, how the agent saves progress, where its data goes and who maintains the runtime. The recommendations below compare documented capabilities and constraints, rather than claiming a measured winner from side-by-side product tests.

Best AI Agent Frameworks and Platforms at a Glance

The frameworks, SDKs and visual builders in this selection have different license terms, including restrictions on some commercial uses. Check the relevant license before redistribution or embedding; hosted services have separate commercial terms.

Tool	Type	Language	Hosting
Vercel AI SDK	Framework	TypeScript	Self-hosted
LangChain and LangGraph	Framework	Python, TypeScript	Self-hosted
CrewAI	Framework	Python	Self-hosted
OpenAI Agents SDK	Vendor SDK	Python, TypeScript	Self-hosted
OpenAI Agents API	Managed harness (beta)	Any (web API)	Hosted (cloud)
Claude Agent SDK	Vendor SDK	Python, TypeScript	Self-hosted
Claude Managed Agents	Managed harness (beta)	Any (web API)	Hosted (cloud)
Google ADK	Vendor SDK	Python, TypeScript, Go, Java, Kotlin	Self-hosted
Microsoft Agent Framework	Vendor SDK	.NET, Python; Go preview	Self-hosted
Cloudflare Agents SDK	Cloud platform	TypeScript	Hosted (cloud)
AWS Bedrock AgentCore	Cloud platform	Any framework	Hosted (cloud)
Dify	Visual builder	No code needed	Both

What Are AI Agent Frameworks and Platforms?

AI agent frameworks and platforms are the tools developers use to build a custom agent that takes a goal, asks a model for the next step, runs it through a tool and repeats. Most of the engineering sits around that loop, in saving progress, retrying failures, limiting access and recording what happened.

Kind	Who Runs the Loop	Examples
Framework	Your code, on your machine or server	Vercel AI SDK, LangGraph, CrewAI
Managed harness	The vendor, on its servers	Claude Managed Agents, OpenAI Agents API
Hosting runtime	Your framework code, on the vendor's servers	AWS Bedrock AgentCore, Cloudflare Durable Objects

Frameworks: Code You Import and Run

An agent framework is a library your application uses to coordinate model calls, tools and state. Model support depends on the framework and provider integration. With a self-managed deployment, your team maintains the application and runtime; connected model services can still be hosted elsewhere.

Platforms: Agents You Call or Host

A hosted platform moves some or all of the work onto the vendor's servers, either running the whole loop or hosting the agent you wrote. Most vendors pair a free library with a paid service, such as LangChain with LangSmith, so read the license and the data terms separately.

When Does Your Team Need an Agent Framework?

An AI agent framework is useful when your team needs to build custom tool use, approvals or state into an application. First check whether an existing product already handles the task. Building your own makes sense when the agent must integrate with product logic or internal systems that the available products cannot support.

Office work belongs on the Business and Enterprise AI Agents page, and fixed repeatable steps suit Workflow Automation Platforms. Agents that act for one person are on Personal AI Agents, and coding agents are in our AI coding tools hub.

Build only when someone on the team will be paged when the agent breaks.

Do You Need a Framework or Custom Code?

A simple model-and-tool loop can be enough for an initial prototype with a small set of actions. Your code sends context to the model, executes permitted tool calls and returns the results until the task finishes. Authentication, retries, persistent state and monitoring still need an implementation; a short loop does not make them disappear.

When Is an Agent Framework Worth Adopting?

A framework is worth adopting once you catch yourself rebuilding durable state, human approval steps, tracing or support for a second model provider by hand. For a smaller task, a short loop may be enough, though a library can still simplify tool definitions and model integration.

Open-Source Frameworks

These three frameworks are free libraries from independent projects, and each works with models from many vendors.

Vercel AI SDK

The Vercel AI SDK is a TypeScript toolkit for AI applications and agents, with model-provider integrations and support for React, Next.js, Vue, Svelte and Node.js. It suits a web team that wants model calls and streaming responses inside an existing application without adopting a separate language stack.

Best for: TypeScript and web teams, especially on React or Next.js.

Main limitation: the library does not remove responsibility for application hosting, persistent state or the model bill. Vercel's managed gateway and workflow services are separate choices; check which services your design actually depends on before assessing portability.

LangChain and LangGraph

LangChain supplies model and tool integrations, while LangGraph provides lower-level orchestration for stateful agents. LangGraph supports durable execution and human review, and it can be used without LangChain. It fits workflows that must branch, pause or resume from saved progress.

Best for: Python teams whose jobs branch, wait for a person or must survive a restart.

Main limitation: saved graph state does not make an external action safe to repeat. Design database writes, payments and messages so retrying a step cannot duplicate the result. That recovery behavior needs testing alongside the graph, especially when an agent resumes after a failure.

CrewAI

CrewAI is a Python framework for agents organized around roles and tasks, such as research and editing. A crew coordinates that team. A Flow is an event-driven workflow that can branch, loop and maintain state, rather than being limited to a fixed sequence.

Best for: Python developers whose work splits naturally into roles.

Main limitation: every extra agent adds model calls, so start with one agent and switch tracing on from the first build.

Model Vendor SDKs

Model makers and cloud vendors publish SDKs with different relationships to their hosted services. Compare the library that runs in your application separately from any managed agent API, because they can have different hosting costs, model support and data-retention rules.

OpenAI Agents SDK and Agents API

The OpenAI Agents SDK is a Python and TypeScript library that runs the agent loop in your application. Your team controls deployment, tools and storage. The separate Agents API runs a managed Codex harness and maintains sessions through OpenAI's service.

Best for: teams that want a small, readable library, or want OpenAI to run long jobs.

Main limitation: the hosted Agents API currently supports US data residency only and does not support Zero Data Retention. Those are API-service restrictions, not a claim that installing the SDK moves your whole application to OpenAI. The SDK's model calls and tracing still need their own data-handling review.

Claude Agent SDK and Claude Managed Agents

The Claude Agent SDK exposes the agent engine behind Claude Code through Python and TypeScript. Claude Managed Agents is the hosted alternative, with server-side sessions and environments. Choose the SDK when you need to manage the runtime yourself; compare Managed Agents when you want Anthropic to operate it.

Best for: teams that like Claude Code's way of working and want it inside their product.

Main limitation: SDK use falls under Anthropic's Commercial Terms, and Managed Agents is currently ineligible for ZDR or HIPAA BAA coverage.

Google ADK

Google's Agent Development Kit provides libraries for building agents and connecting them to tools. Its language options include Python, TypeScript, Go, Java and Kotlin, with capabilities that vary by implementation. Check the documentation for your chosen language and deployment target instead of assuming every example works across all SDKs.

Best for: Java and Go teams, and anyone hosting on Google Cloud.

Main limitation: language implementations and model integrations do not necessarily support the same features. Verify state handling, tool support and deployment requirements in the specific SDK you plan to use before committing to it.

Microsoft Agent Framework

Microsoft Agent Framework brings Microsoft's agent and workflow tooling into a shared framework, with migration guidance from Semantic Kernel and AutoGen. It supports .NET and Python, and now offers a Go public preview. The Go implementation has documented feature gaps, so compare capabilities by language.

Best for: C# and .NET teams, on Azure or elsewhere.

Main limitation: migration from existing Semantic Kernel or AutoGen code requires checking the new APIs and behavior. Pin dependencies and test tool calls, state and approvals on your chosen version; a familiar Microsoft model provider does not make an older example compatible.

Cloud Agent Platforms

Cloud agent platforms provide hosting, state storage, identity and monitoring around your agent. They can reduce infrastructure work, but moving away may require replacing those services and exporting stored sessions. The framework's license and the hosting service's commercial terms answer different questions.

Cloudflare Agents SDK, Durable Objects and MCP Hosting

Cloudflare Agents runs durable agents on Durable Objects, with per-agent SQL storage, connections and scheduled work. The SDK supports tools and MCP integrations alongside that runtime. Its billing includes compute and storage; estimating cost from Workers CPU time alone misses part of the bill.

Best for: TypeScript teams that need per-agent state on the edge.

Main limitation: each agent's state lives on Cloudflare's network, so export it on a schedule.

AWS Bedrock AgentCore

Amazon Bedrock AgentCore is AWS's set of services for running agents built with any framework, including CrewAI, LangGraph and the OpenAI Agents SDK. Its pieces cover a per-session runtime, a managed harness, a gateway that turns APIs into MCP tools, and memory, identity and tracing.

Best for: companies that already run their workloads on AWS.

Main limitation: it runs only on AWS, so it is a commitment for anyone else.

Visual Builders You Can Self-Host

Visual agent builders let you connect model calls, data sources and actions in a canvas. A self-hosted option can help a team control its deployment, but it still needs credentials, updates and monitoring. Choose a visual builder when its available components cover the process you need to maintain.

Dify

Dify is a visual builder for agents, chat apps and workflows with document search built in, available as Dify Cloud or a free self-hosted Community Edition. Its license bars running it as a multi-tenant service or removing the Dify logo without permission.

Best for: teams that want a visual builder on a private network.

Main limitation: an exposed builder is a code-execution service, so keep it off the public internet.

Other Frameworks Worth Knowing

These options address additional needs, from document retrieval to visual configuration and agents inside chat tools. Compare the specific capability with your application before adding another framework to the stack.

Tool	What It Is	Hosting
Langflow	An MIT visual builder where each box is editable Python	Self-hosted
LlamaIndex	Agents that work over your own documents	Both
smolagents	Hugging Face's Python library for tool-using agents	Self-hosted
CopilotKit OpenTag	An open Slack and Teams agent built on LangGraph	Both

Patch Langflow promptly, because CVE-2026-33017 allowed code execution without a login.

Where Does MCP Fit?

MCP, the Model Context Protocol, is the open standard that lets an agent use outside tools and data, and it connects tools without running agents. The major frameworks here support it, and our guide to MCP servers and hooks in Claude Code covers connecting one safely.

How Are Agent Frameworks and Platforms Priced?

Many agent libraries have no separate software fee, while model inference, hosting and optional managed services are billed separately. LangSmith Plus starts at US$39 per seat per month plus usage; Claude Managed Agents charges model usage plus US$0.08 per active session-hour. The table distinguishes library access from those service charges.

Tool	Free Option	Software and Service Costs
Vercel AI SDK	Library available without a separate fee	No fee for the SDK
LangChain and LangGraph	Yes, plus a free LangSmith seat	LangSmith from $39 a seat a month
CrewAI	Free library; AMP Basic includes 50 workflow runs a month	Enterprise by quote
OpenAI Agents SDK and Agents API	The SDK	Standard usage rates
Claude Agent SDK and Managed Agents	SDK access; model usage is billed	SDK: model usage; Managed Agents: usage plus US$0.08 per active session-hour
Google ADK	ADK library; hosting billed separately	Agent Runtime by the hour
Microsoft Agent Framework	Library available without a separate fee	No fee for the framework
Cloudflare Agents SDK	Workers Free includes Durable Objects	Workers Paid from $5 a month
AWS Bedrock AgentCore	Free-tier eligibility depends on the account	Pay as you go
Dify	Community Edition and Sandbox plan	From $590 a year per workspace

Set a step limit and a spending alert before any agent's first unattended run, because a looping agent keeps calling the model after it stops making progress.

How Should You Choose an Agent Framework?

Choose an agent framework by the language your team maintains, the work the agent must perform and the way it recovers from failure. Prototype one representative task, including an interrupted run, before comparing these requirements:

• Saved state: jobs that run for hours or wait for a person need resume support.
• Tracing: capture enough model and tool activity to explain failures, with secrets redacted and sensitive data handled under your retention policy.
• Portability: changing a model or host can require changes to tools, state and evaluation. Check how sessions and data can be exported.
• Data handling: check where sessions and traces are stored. OpenAI Agents API is US-only, and both it and Claude Managed Agents currently exclude Zero Data Retention.
• Security: agents that read files or web pages are open to prompt injection, which our application security hub covers.

Compare candidates on the same small set of representative tasks. Include a failed tool call, a restart after a completed write, and a request that requires human approval. Check whether the task finishes correctly, whether an action is duplicated, and whether the trace explains what happened. Record model and infrastructure costs separately so a successful demo does not hide an expensive runtime.

Which Agent Framework Should You Start With?

For a TypeScript web application, start by evaluating Vercel AI SDK; for branching work with saved state, evaluate LangGraph; for an existing .NET stack, evaluate Microsoft Agent Framework. Run one representative task locally, then compare deployment options. A team's language narrows the choice, but recovery and integration requirements decide it.

Your Team	Start With
TypeScript web team	Vercel AI SDK
Python team	LangGraph, without LangChain at first
.NET shop	Microsoft Agent Framework
Already on AWS	Your framework on AgentCore Runtime
Wants a visual builder	Dify, self-hosted on a private network

Choose CrewAI when distinct agent roles help organize the task, and compare that design with a simpler single-agent version. Hosted agent services are useful when you want managed sessions and less runtime maintenance, provided their beta status and data-retention terms meet your requirements.

What Else Do People Ask?

Are AI agent frameworks free?

Many AI agent SDKs have no separate software license fee, but model calls, hosting and managed services can still cost money. Product terms also differ: Dify restricts some commercial deployments, while the Claude Agent SDK is subject to Anthropic's commercial terms. A free library does not make a running agent free.

Is LangChain still worth learning?

LangChain is useful when its model integrations and agent abstractions match your Python or TypeScript application. LangGraph is the lower-level option for explicit state, branching and recovery, and can be used independently. Choose the layer your task needs rather than learning both before building a simple prototype.

Is the Vercel AI SDK only for Vercel?

The Vercel AI SDK can be used outside Vercel hosting, including in Node.js applications and supported web frameworks. The SDK supplies model integrations and application features; Vercel's gateway, hosting and workflow services are separate products. Check your dependencies to see whether your application uses any of those services.

Can I use Claude models with the OpenAI Agents SDK?

The OpenAI Agents SDK supports model-provider integrations, including routes to models outside OpenAI. Using Claude requires a compatible provider integration and separate Anthropic credentials and billing. Check which tools and features that integration supports; the Claude Agent SDK instead builds around Claude Code's agent engine.

What happened to AutoGen and Semantic Kernel?

Microsoft Agent Framework is the successor path for agent development using Semantic Kernel and AutoGen. Microsoft provides migration guidance for both projects. Existing applications still need compatibility checks, so follow the migration guide for the APIs you use instead of treating a package-name change as a complete migration.

Can AI agent frameworks use local models?

AI agent frameworks can use local models when a supported provider integration connects to the local server and the model supports the required behavior, such as tool calling. That support varies by SDK and model. Local inference does not prevent tools, MCP connections or tracing from sending data elsewhere.

Where to Go Next

For repeatable jobs across existing apps, compare workflow automation platforms before building an agent from scratch.

The other category pages in the AI Agents and Automations cluster cover agents you can use without building anything, from business agents like Claude Tag to personal agents like OpenClaw. For an open-source agent to run yourself today, start with our OpenCode guide.

Product documentation and pricing were reviewed on October 29, 2026. Recheck plan terms and preview features before deploying an agent.