Frameworks for building AI agents and multi-agent systems
Open-source LLM application development platform. Visual workflow builder for creating AI agents, chatbots, and RAG applications. 🔗 Stars 50k+ github.com/langgenius/dify
Microsoft's open-source multi-agent conversation framework. Enables building LLM applications with multiple agents that converse and collaborate. 🔗 Stars 35k+ github.com/microsoft/autogen
轻量级多智能体编排框架。可用最少代码定义智能体角色、分配任务并运行协作式AI团队。 🔗 文章：/post/best-python-ai-agent-frameworks-2026 | Stars 25k+
最流行的开源框架，用于构建基于LLM的应用。支持智能体、RAG、工具调用和多步骤工作流。 🔗 文章：/post/best-python-ai-agent-frameworks-2026 | Stars 100k+
IBM Research 开源的智能体框架，围绕互操作协议构建，多智能体编排与可复用技能，提供 Python/TypeScript 版本。
字节跳动开源的深度研究超级智能体框架，编排子智能体、持久记忆与沙箱执行完成长时任务，2026 年初登顶 GitHub Trending。
开源通用 AI 智能体框架（灵感来自 Manus），无需邀请码本地运行，支持规划、工具调用、浏览器操作与代码执行。
Google 开源的智能体开发套件（Agent Development Kit），Python/TypeScript/Java/Go 四语言 SDK 对齐，支持 A2A 协议与 Google Cloud 深度集成。
Kotaemon 是 Cinnamon 开源的 RAG 文档问答界面，支持本地大模型与云端模型，提供引用溯源与知识库管理。
Mirascope 是一个专注于开发体验的 LLM 应用框架，为 Python 开发者提供简洁抽象，统一调用多家大模型 API。
Hermes Agent 是 Nous Research 出品的开源通用 AI 智能体框架：本地优先、可自我改进，支持工具调用、浏览器自动化、后台任务与多智能体协作，是 2026 年增长最快的开源智能体框架之一。 🔗 文章: https://ainomam.com/post/hermes-agent-数据库瘦身更新-2dba16 | GitHub Stars: 236k+
NVIDIA NeMo Agent Toolkit 是英伟达开源的企业级智能体开发工具包，覆盖智能体组合、工具调用、编排与部署模式，适合已在 NVIDIA 生态或需要企业级智能体基础设施的团队。 🔗 文章: https://ainomam.com/post/ai-agent-monitoring-with-datadog-complete-setup-guide-for-2026-20260723 | GitHub Stars: 2.6k+
GPTeam 是 MIT 许可的 Python 多智能体协作框架，让多个角色型 AI 智能体在同一环境里互动、协作完成目标，适合研究角色分工、多智能体通信与群体行为实验。 🔗 文章: https://ainomam.com/post/ai-handshake-part-6 | GitHub Stars: 1.7k+
VoltAgent 是一个开源 TypeScript AI Agent 工程平台，内置记忆、RAG、护栏、工具调用、MCP 支持、工作流与可观测性，把智能体编排做成企业级工程，MIT 许可。 🔗 文章: https://ainomam.com/post/ai-agent-monitoring-with-github-track-agent-performance-in-2026-20260723 | GitHub Stars: 10k+
Stagehand 是 Browserbase 出品的浏览器自动化框架，用自然语言驱动 Playwright 操作网页：点按钮、填表单、抓数据，一行指令代替一堆选择器代码。适合做网页 AI 智能体和爬虫任务，支持 TypeScript/Python。 🔗 GitHub: https://github.com/browserbase/stagehand | ⭐ 24k+
Zep 是企业级 AI Agent 记忆基础设施，通过上下文图谱为 Agent 提供亚 200ms 检索的企业级记忆，SOC 2 合规，被 S&P 等机构验证。
An open-source AI agent for browser automation that handles complex workflows on the web. Skyvern uses LLM + computer vision to navigate websites, fill forms and extract data, even when page structures change. Alternative to brittle DOM-selector-based scraping.
Coinbase's open-source toolkit for building AI agents that can interact with crypto and blockchain. Provides tools for payments, wallet management, onchain transactions and DeFi operations, letting LLM agents execute real onchain actions with CDP integration.
Microsoft's open-source framework for building production-grade AI agents. Provides agent runtime, workflow orchestration (Workflows and Semantic Kernel-based agents), multi-agent collaboration, and integration with Azure AI Foundry. Successor of Semantic Kernel for agent-centric development.
微软开源的受控生成框架，用模板和约束控制 LLM 输出结构，显著降低 token 消耗与推理成本，让结构化输出更可靠。
Anthropic 官方发布的 Claude 智能体开发 SDK，可让 Claude 在沙箱中自主运行代码、读写文件、调用工具，支持 TypeScript 与 Python。
Google 官方多智能体开发框架（Agent Development Kit），Python 编写，与 Gemini/Vertex AI 深度集成。
Microsoft's generalist multi-agent system — an Orchestrator agent coordinates web, coding and file agents to complete complex tasks.
Best official SDK for building AI agents — lightweight multi-agent orchestration with tools and handoffs.
Best framework for programming LLMs — optimize prompts and pipelines like a compiler instead of hand-tuning.
Best visual AI app builder — drag-and-drop RAG, agents and multi-model workflows, exportable as API.
Best open-source library that lets AI agents control a real browser — click, type and extract via natural language.
Pydantic AI is a type-safe Python agent framework — model-agnostic, with structured outputs guaranteed by Pydantic validation.
Microsoft AutoGen is a multi-agent conversation framework — build AI agents that collaborate, delegate, and solve tasks together through structured conversations.
Durable functions and event-driven workflow orchestration with built-in queues, retries, and observability
OpenAI-compatible LLM gateway/proxy that unifies 100+ models behind one API with budgets and guardrails
Best open-source React framework to embed AI copilots - add chat, agents and automation to any web app with ready-made UI components and backend orchestration.
Best open-source framework for building AI agents - create agents with memory, knowledge bases and tool calling in Python, deploy to AWS/GCP, with built-in monitoring and evaluation.
Best open-source AI framework from Google for production apps - unified GenAI flows with built-in observability, evaluation, and tracing; TypeScript and Go support.
Best TypeScript toolkit for building AI apps - unified API for LLM streaming, tool calling, and agents across 100+ providers; works with React, Next.js, Node, and serverless; open source.
Best open-source RAG engine for accurate document Q&A - deep document understanding with template-free chunking, citation-grounded answers, and a visual agent builder, deployed easily via Docker.
LLM的结构化生成——引导模型输出有效JSON、正则表达式和类型安全数据，语法有保证。该库为可靠的智能体工具调用提供支持。GitHub 15K+星标。
Stanford NLP 的 LLM 编程框架——像软件一样声明、优化和编译提示词。将提示工程转化为可测试、自我改进的代码。GitHub 星标 36K+。
开源实时语音智能体框架：将LLM、STT、TTS与WebRTC整合为可对话的AI助手。GitHub星标超13k。
阿里ModelScope开源多智能体框架，支持消息传递、分布式调度、可视化调试及MCP。28k+星标。
Hugging Face 的极简智能体框架，「代码即动作」哲学，轻量易学。
最佳官方Python框架，用于智能体工作流——智能体、交接、护栏、MCP支持及追踪。
LlamaIndex是一个用于构建LLM应用的数据框架，尤其适用于基于文档的AI智能体。它最初是一个解析库，后来发展为企业文档智能体底下的数据层。提供先进的索引、检索、解析和智能体RAG能力，深受生产级文档工作流信赖。
Mastra是一个TypeScript优先的开源框架，用于构建AI智能体，由Gatsby团队打造。它提供工作流、RAG管道、记忆管理、MCP支持和内置评估——全部基于TypeScript。获YC支持，融资1300万美元。专为偏好Node.js生态的工程团队设计。
MetaGPT是一个开源的多智能体框架，模拟一家软件公司，其中智能体扮演产品经理、架构师、工程师和QA的角色。
TaskWeaver 是微软开源框架，用于构建有状态、代码优先的AI智能体，能够规划并执行复杂任务。
Flowise 是一款开源低代码工具，用于可视化构建 LLM 应用和 AI 智能体。通过拖拽界面创建 RAG 聊天机器人、智能体工作流和 API 端点，无需编写代码。
开源多智能体编排框架，用于构建、部署和扩展AI智能体集群。支持层级智能体结构、工具调用和复杂工作流自动化。 🔗 GitHub: github.com/kyegomez/swarms | Stars: 15k+
Haystack是一个开源NLP框架，用于构建生产级搜索和问答系统。支持LLM、嵌入、向量数据库和检索增强生成（RAG）。基于Python，采用模块化管道设计。
Autonomous AI agent framework backed by a16z. Designed for creating personality-driven agents with memory and tool use. 🔗 Stars 10k+ github.com/elizaOS/eliza
Multi-modal agent framework (formerly Phidata). Build agents that can see, hear, and interact with any data source. 🔗 Stars 15k+ github.com/agno-agi/agno
Agent framework by the Pydantic team. Type-safe, structured output guaranteed — built for production-grade AI agents. 🔗 Stars 5k+ github.com/pydantic/pydantic-ai
Microsoft's lightweight AI orchestration SDK for .NET and Python. Integrates LLMs with traditional programming for agentic workflows. 🔗 Microsoft SDK: learn.microsoft.com/semantic-kernel