Agent frameworks

An AI agent framework is a library that supplies the building blocks of an agent: a loop that calls a model, executes the tool calls it returns, manages state and memory, and coordinates multiple agents or workflow steps. A framework saves teams from writing that orchestration code from scratch.

The frameworks differ mainly in abstraction level and language. LangGraph, Microsoft Agent Framework, and Eino expose graphs or workflows with explicit state, CrewAI and AG2 organize teams of agents, and the OpenAI Agents SDK, Pydantic AI, and Strands Agents keep the core loop small. Agent toolkits such as the Claude Agent SDK, Deep Agents, and Letta ship an assembled agent with built-in tools.

Most entries are open source and provider-agnostic, and most connect to tools through the Model Context Protocol. Python has the widest choice, TypeScript teams can use Mastra, VoltAgent, or the Vercel AI SDK, and JVM, Go, and Rust developers have Spring AI, LangChain4j, Koog, Eino, and Rig. Microsoft has archived AutoGen and points Semantic Kernel users to Microsoft Agent Framework.

Agent frameworks compared

PlatformOpen sourceDeploymentBest for
AG2Yes (Apache-2.0)LibraryTeams continuing AutoGen-style multi-agent patterns
AgnoYes (Apache-2.0)Library, Self-hostedRunning agents as services in your own cloud
BeeAI FrameworkYes (Apache-2.0)LibraryPython or TypeScript agents under open governance
CAMELYes (Apache-2.0)LibraryMulti-agent research and synthetic data generation
Claude Agent SDKNoLibraryAgents that need file, shell, and code tools
Cloudflare Agents SDKYes (MIT)Library, Managed cloudStateful agents hosted on Cloudflare
CrewAIYes (MIT)LibraryRole-based multi-agent collaboration in Python
Deep AgentsYes (MIT)LibraryLong-horizon tasks that need subagents and a filesystem
DSPyYes (MIT)LibraryOptimizing LLM programs against measurable metrics
EinoYes (Apache-2.0)LibraryGo services that need agents and workflows
GenkitYes (Apache-2.0)LibraryApp developers adding AI features across several languages
Google Agent Development Kit (ADK)Yes (Apache-2.0)LibraryAgents deployed on Google Cloud
HaystackYes (Apache-2.0)LibraryProduction RAG and agent pipelines in Python
KoogYes (Apache-2.0)LibraryKotlin and JVM teams building agents
LangChainYes (MIT)LibraryAssembling agents from prebuilt integrations
LangChain4jYes (Apache-2.0)LibraryJava teams that want LangChain-style building blocks
LangGraphYes (MIT)LibraryLong-running agent workflows that need explicit control
LettaYes (Apache-2.0)CLI, Self-hosted, Managed cloud, LibraryLong-lived agents that keep and edit their own memory
LlamaIndexYes (MIT)LibraryAgents that work over documents and enterprise data
MastraYes (Apache-2.0)LibraryTypeScript teams building agents into web apps
Microsoft Agent FrameworkYes (MIT)LibraryPython and .NET teams building on Azure
OpenAI Agents SDKYes (MIT)LibraryMinimal abstractions for multi-agent handoffs
Pydantic AIYes (MIT)LibraryType-safe Python agents with validated structured output
RigYes (MIT)LibraryRust services that call LLMs and tools
Semantic KernelYes (MIT)LibraryMaintaining existing Semantic Kernel applications
smolagentsYes (Apache-2.0)LibraryMinimal code-writing agents and experiments
Spring AIYes (Apache-2.0)LibrarySpring Boot teams adding agents to Java services
Strands AgentsYes (Apache-2.0)LibraryModel-driven agents on AWS or any other cloud
Vercel AI SDKYes (Apache-2.0)LibraryTypeScript web apps that stream model output to users
VoltAgentYes (MIT)LibraryTypeScript agents with built-in observability

Profiles

  • AG2 logo

    AG2

    ag2.ai

    AG2 is an open-source Python framework for multi-agent systems from the creators of AutoGen, maintained by volunteers from several organizations; version 1.0 introduced a protocol-driven package and moved the classic ConversableAgent and GroupChat API to AG2 Classic.

    • Open source · Apache-2.0
    • Library
    • Python
  • Agno logo

    Agno

    agno.com

    Agno is an open-source Python framework and agent platform: developers build agents, teams, and workflows with the Agno SDK and run them as durable services on AgentOS, which adds APIs, MCP, authentication, approvals, and scheduling in their own cloud.

    • Open source · Apache-2.0
    • Library
    • Self-hosted
    • Python
  • BeeAI Framework

    framework.beeai.dev

    BeeAI Framework is an open-source framework for building multi-agent systems in Python and TypeScript, first released by IBM developers and now part of the Linux Foundation AI & Data program, with workflows and MCP and A2A integrations.

    • Open source · Apache-2.0
    • Library
    • Python
    • TypeScript
  • CAMEL logo

    CAMEL

    camel-ai.org

    CAMEL is an open-source Python multi-agent framework from the CAMEL-AI community, built to study how agent systems scale, with components for synthetic data generation, world simulation, and task automation by cooperating agents.

    • Open source · Apache-2.0
    • Library
    • Python
  • Claude Agent SDK logo

    Claude Agent SDK

    Anthropic · code.claude.com

    Anthropic's Claude Agent SDK exposes the agent loop, built-in tools, context management, and permission system behind Claude Code as a library in Python and TypeScript for building agents that edit files, run commands, and use MCP servers.

    • Library
    • Python
    • TypeScript
  • Cloudflare Agents SDK logo

    Cloudflare Agents SDK

    Cloudflare · developers.cloudflare.com

    Cloudflare's Agents SDK is an open-source TypeScript framework for agents that run on Cloudflare Workers, where each agent lives in a Durable Object with its own state, WebSocket connections, scheduling, and tools, including remote MCP servers.

    • Open source · MIT
    • Library
    • Managed cloud
    • TypeScript
  • CrewAI

    crewai.com

    CrewAI is an open-source Python framework for multi-agent workflows that combines Crews, groups of role-based agents that collaborate on a task, with Flows for event-driven control, plus a commercial control plane for deployment and monitoring.

    • Open source · MIT
    • Library
    • Python
  • Deep Agents logo

    Deep Agents

    LangChain · docs.langchain.com

    Deep Agents is LangChain's open-source agent toolkit built on LangGraph, with planning, subagents that run in isolated context windows, a pluggable filesystem, shell access, persistent memory, and human approval of tool calls for long-horizon tasks.

    • Open source · MIT
    • Library
    • Python
    • TypeScript
  • DSPy logo

    DSPy

    dspy.ai

    DSPy is an open-source Python framework for programming language models with declarative modules and signatures instead of hand-written prompts, with optimizers that tune prompts and weights against a metric for agents, RAG pipelines, and classifiers.

    • Open source · MIT
    • Library
    • Python
  • Eino logo

    Eino

    ByteDance · cloudwego.io

    Eino is an open-source Go framework for LLM applications from ByteDance's CloudWeGo project, with reusable components, an Agent Development Kit for tool use, multi-agent coordination, and human-in-the-loop interrupts, and graph composition that follows Go conventions.

    • Open source · Apache-2.0
    • Library
    • Go
  • Genkit logo

    Genkit

    Google · genkit.dev

    Genkit is Google's open-source framework for building AI-powered apps and agents in TypeScript, Go, Python, and Dart, with one interface to models from Gemini, Claude, OpenAI, and other providers; it began as Firebase Genkit.

    • Open source · Apache-2.0
    • Library
    • TypeScript
    • Go
    • Python
  • Google Agent Development Kit (ADK) logo

    Google Agent Development Kit (ADK)

    Google · adk.dev

    Google's Agent Development Kit is an open-source, code-first framework for building, evaluating, and deploying agents in Python, TypeScript, Go, Java, and Kotlin, with built-in Gemini access and adapters for other model providers.

    • Open source · Apache-2.0
    • Library
    • Python
    • TypeScript
    • Go
  • Haystack

    deepset · haystack.deepset.ai

    Haystack is deepset's open-source Python framework for building agents and RAG systems from modular components connected in pipelines, with integrations for model providers and document stores and a commercial Haystack Enterprise Platform for deployment.

    • Open source · Apache-2.0
    • Library
    • Python
  • Koog logo

    Koog

    JetBrains · jetbrains.com

    Koog is JetBrains' open-source Kotlin framework for building AI agents that run on the JVM, Android, iOS, JavaScript, and WebAssembly, aimed at teams that want agents written in idiomatic Kotlin across platforms.

    • Open source · Apache-2.0
    • Library
    • Kotlin
  • LangChain

    langchain.com

    LangChain is an open-source framework for building agents from prebuilt agent architectures and a large library of integrations with model providers, vector stores, retrievers, and tools, with LangGraph underneath for lower-level orchestration.

    • Open source · MIT
    • Library
    • Python
    • TypeScript
  • LangChain4j logo

    LangChain4j

    docs.langchain4j.dev

    LangChain4j is an open-source Java library for building LLM applications and agents on the JVM, with unified APIs across model providers and embedding stores plus components for tool calling, chat memory, agents, and retrieval-augmented generation.

    • Open source · Apache-2.0
    • Library
    • Java
  • LangGraph

    LangChain · langchain.com

    LangGraph is an open-source, low-level orchestration framework from LangChain for building stateful agents as graphs, with durable execution, persistence, streaming, and human-in-the-loop checkpoints, in Python and TypeScript.

    • Open source · MIT
    • Library
    • Python
    • TypeScript
  • Letta logo

    Letta

    letta.com

    Letta, formerly MemGPT, builds stateful agents with long-term memory; its open-source Letta Code agent lets agents rewrite their own memory, skills, and prompts, and runs through a CLI, a self-hosted server, Letta Cloud, or the Letta Agent SDK.

    • Open source · Apache-2.0
    • CLI
    • Self-hosted
    • Managed cloud
    • Library
    • TypeScript
  • LlamaIndex logo

    LlamaIndex

    llamaindex.ai · San Francisco, United States · founded 2023

    LlamaIndex is an open-source Python framework for agents and multi-step workflows over private data, with connectors, indexing, and retrieval components; the company now focuses mainly on LlamaParse, and its TypeScript port is deprecated.

    • Open source · MIT
    • Library
    • Python
  • Mastra logo

    Mastra

    Kepler Software · mastra.ai

    Mastra is an open-source TypeScript framework for agents, graph-based workflows, memory, RAG, evals, and observability, with human-in-the-loop suspend and resume and a local Studio for testing; enterprise directories use a separate license.

    • Open source · Apache-2.0
    • Library
    • TypeScript
  • Microsoft Agent Framework logo

    Microsoft Agent Framework

    Microsoft · learn.microsoft.com

    Microsoft Agent Framework is an open-source SDK for Python and .NET, generally available since April 2026, that succeeds Semantic Kernel and the archived AutoGen for building agents and multi-agent workflows, with support for MCP and A2A.

    • Open source · MIT
    • Library
    • Python
    • C#
  • OpenAI Agents SDK logo

    OpenAI Agents SDK

    OpenAI · openai.github.io

    The OpenAI Agents SDK is a lightweight open-source framework for Python and TypeScript built on agents, handoffs between agents, guardrails, sessions, and built-in tracing, and it also works with models from other providers.

    • Open source · MIT
    • Library
    • Python
    • TypeScript
  • Pydantic AI logo

    Pydantic AI

    Pydantic · pydantic.dev

    Pydantic AI is an open-source Python agent framework from the Pydantic team with a typed agent loop, validated structured outputs and tool arguments, model fallbacks across many providers, and durable execution through integrations such as Temporal.

    • Open source · MIT
    • Library
    • Python
  • Rig

    rig.rs

    Rig is an open-source Rust library for building LLM applications and agents, with one API across many model providers, type-safe tools, structured output, and integrations with vector stores for retrieval.

    • Open source · MIT
    • Library
    • Rust
  • Semantic Kernel logo

    Semantic Kernel

    Microsoft · learn.microsoft.com

    Semantic Kernel is Microsoft's open-source, model-agnostic SDK for building agents and multi-agent systems in Python, .NET, and Java, with plugins, vector store connectors, and MCP support; Microsoft now directs new projects to Microsoft Agent Framework.

    Microsoft names Microsoft Agent Framework as the successor and continues to fix critical bugs and security issues in Semantic Kernel v1.x.

    • Open source · MIT
    • Library
    • Python
    • C#
    • Java
  • smolagents logo

    smolagents

    Hugging Face · huggingface.co

    smolagents is Hugging Face's open-source Python library for running agents in a few lines of code, with deliberately minimal abstractions and a focus on code agents that write their actions as Python, plus support for models from many providers.

    • Open source · Apache-2.0
    • Library
    • Python
  • Spring AI

    Broadcom · spring.io

    Spring AI is the Spring project's open-source framework for adding AI to Java applications, with a ChatClient API, advisors, tool calling, structured output, chat memory, and MCP client and server starters that follow Spring Boot conventions.

    • Open source · Apache-2.0
    • Library
    • Java
  • Strands Agents logo

    Strands Agents

    Amazon Web Services · strandsagents.com

    Strands Agents is an open-source SDK from AWS for building agents in Python and TypeScript that run in-process without a hosted control plane, with model portability across Bedrock, Anthropic, OpenAI, and Gemini, MCP, multi-agent patterns, and a prebuilt agent loop.

    • Open source · Apache-2.0
    • Library
    • Python
    • TypeScript
  • Vercel AI SDK logo

    Vercel AI SDK

    Vercel · ai-sdk.dev

    The AI SDK from Vercel is an open-source TypeScript toolkit for AI applications and agents, with one interface across model providers, streaming, fallbacks, structured output, a ToolLoopAgent class for tool-calling loops, and chat interface support.

    • Open source · Apache-2.0
    • Library
    • TypeScript
  • VoltAgent logo

    VoltAgent

    voltagent.dev

    VoltAgent is an open-source TypeScript agent framework with typed tools, memory adapters, a workflow engine, supervisor and sub-agent coordination, MCP support, and resumable streaming, paired with the VoltOps console for observability, evals, and deployment.

    • Open source · MIT
    • Library
    • TypeScript

How to choose

01Abstraction level
Low-level graph or workflow primitives give control over every step, while higher-level agent classes and prebuilt agents get a prototype running faster.
02Language support
Python, TypeScript, .NET, Java, Kotlin, Go, or Rust support should match the team's production stack.
03Model support
Provider-agnostic frameworks let you switch models or mix providers without rewriting agent code.
04State and durability
Checkpointing, persistence, and human-in-the-loop pauses matter for long-running or approval-gated tasks.
05Multi-agent and protocols
Handoffs, supervisor patterns, and support for MCP and the A2A protocol determine how agents and tools connect.
06Observability and deployment
Built-in tracing, OpenTelemetry export, and a clear path to production hosting reduce operational work.

Frequently asked questions

What is an AI agent framework?

An AI agent framework is a software library for building applications in which a language model's output drives a sequence of tool calls to complete a task. It handles the model and tool loop, conversation state, memory, error handling, and multi-agent coordination, so developers write tools and instructions rather than orchestration code. LangGraph, CrewAI, and the OpenAI Agents SDK are examples.

LangGraph vs CrewAI: what is the difference?

LangGraph is a low-level framework that models an agent as a graph of nodes and edges with explicit shared state, checkpoints, and human-in-the-loop interrupts, which gives fine control over each step. CrewAI organizes work around Crews of role-based agents that collaborate on tasks, plus Flows for event-driven control. Both are MIT-licensed; CrewAI is Python-only, while LangGraph also has a TypeScript version.

Which AI agent framework does Microsoft recommend?

Microsoft recommends Microsoft Agent Framework, an open-source SDK for Python and .NET that reached general availability in April 2026. It succeeds AutoGen, whose repository is now marked archived and no longer maintained, and Semantic Kernel, which still receives critical bug and security fixes while most new features go to Agent Framework. Microsoft publishes migration guides from both projects.

Which AI agent framework is best for production?

No single framework is best for every production system. Teams usually decide on language first, then on how much control they need: graph frameworks such as LangGraph suit long-running, approval-gated workflows, while smaller loops such as Pydantic AI or the OpenAI Agents SDK suit focused tools. Check durability, tracing, and model portability, and build the same small task in two candidates before standardizing.

Related guides

AI agents hub

Other categories

Product details change quickly; check each vendor’s site for current features, licensing and pricing.