Agent Frameworks & SDKs

Anthropic's production-grade agent runtime with deep MCP integration, computer use capabilities, and a developer-first design philosophy. It provides an agentic loop that automatically handles tool calls, supports multi-turn conversations, and includes built-in guardrails for safe agent behavior.

• - Production agent systems with tool use
• - Computer use automation (browser, desktop)

OpenAI's official SDK for building production-ready agents with managed infrastructure, tool use, handoffs between specialized agents, and built-in guardrails. It provides a minimal yet powerful set of primitives: Agents, Handoffs, Guardrails, and a Runner to orchestrate everything.

• - Production AI assistants with handoffs
• - Tool-augmented chatbots and copilots

Google's comprehensive Agent Development Kit for building multi-agent systems powered by Gemini and other models. It provides a layered architecture supporting simple LLM agents, pipeline agents with sequential/parallel/loop workflows, and custom agents with arbitrary orchestration logic. Deep integration with Google Cloud services and Vertex AI.

• - Gemini-powered multi-modal agents
• - Google Cloud-integrated enterprise workflows

Microsoft's framework for building multi-agent systems where agents can converse with each other, use tools, and collaborate on complex tasks. AutoGen 0.4 introduced a complete rewrite with an event-driven architecture, async-first design, and a modular component system. Complements Semantic Kernel for enterprise scenarios with Azure integration.

• - Multi-agent debates and discussions
• - Enterprise workflows with Azure integration

AWS's open-source SDK for building AI agents that integrates seamlessly with Amazon Bedrock and other AWS services. Strands follows a model-driven approach where the AI model acts as the orchestrator, deciding which tools to call and when. It supports any model provider while providing first-class Bedrock integration for production deployments.

• - AWS-native AI agent applications
• - Amazon Bedrock-powered workflows

Graph-based framework for building stateful, multi-actor LLM applications with cycles, controllability, and persistence. LangGraph models agent workflows as state machines where nodes are functions and edges define transitions, including conditional routing. It provides built-in checkpointing, human-in-the-loop support, and seamless LangSmith integration for debugging.

• - Complex multi-step agent workflows with branching logic
• - Human-in-the-loop approval and review systems

Role-based multi-agent framework where agents have defined roles, backstories, and goals. CrewAI emphasizes natural collaboration and delegation between agents organized as a 'crew'. It provides a high-level abstraction that makes it easy to create teams of AI agents that work together on complex tasks, with built-in support for sequential and hierarchical processes.

• - Business process automation with specialized teams
• - Content creation pipelines (research, write, edit)

Community-driven evolution of the original AutoGen project, focused on open multi-agent conversation patterns and group decision-making. AG2 emphasizes conversable agents that can engage in flexible, dynamic conversations with each other, supporting human participation and complex group chat scenarios.

• - Multi-agent conversations and debates
• - Group decision-making with diverse agent perspectives

Document-centric agent framework built from industry-leading RAG foundations. LlamaIndex focuses on structured data ingestion, indexing, and agentic document workflows. Its agent abstraction combines reasoning with data retrieval, making it the go-to framework for knowledge-intensive applications that need to reason over private data.

• - Knowledge management and document Q&A systems
• - Agentic RAG with multi-step retrieval

Ultra-minimal agent framework from Hugging Face where agents write and execute Python code to achieve goals. Instead of traditional tool calling via JSON, smolagents uses a 'code agent' approach where the LLM generates Python code snippets that are executed in a sandboxed environment. This leads to more flexible and composable agent behavior.

• - Learning and understanding agent fundamentals
• - Code-generation and data analysis tasks

Type-safe agent framework built by the creators of Pydantic. PydanticAI leverages Python's type system for structured agent outputs, dependency injection, and model-agnostic design. It brings the reliability and developer experience of Pydantic validation to the world of AI agents, ensuring outputs conform to defined schemas.

• - Type-safe agent outputs with guaranteed schema compliance
• - Structured data extraction from unstructured text

Lightweight, model-agnostic agent framework focused on simplicity, speed, and multi-modal support. Formerly known as Phidata, Agno provides a clean API for building agents with tools, knowledge bases, memory, and team collaboration. It supports text, image, audio, and video modalities with a unified interface.

• - Fast agent prototyping with minimal boilerplate
• - Multi-modal agents (text, image, audio, video)

Production-ready framework for building composable NLP and AI pipelines. Haystack 2.x uses a component-based architecture where each component (generators, retrievers, converters, etc.) is a self-contained unit that can be connected into pipelines. While primarily a pipeline framework, it supports agentic patterns through its ChatGenerator components and tool use capabilities.

• - Production NLP and RAG pipelines
• - Document processing and indexing at scale

TypeScript-first agent framework with built-in workflows, RAG, integrations, and evaluation tools. Mastra provides a cohesive toolkit for building production agents in the TypeScript/Node.js ecosystem, with first-class support for structured outputs, tool calling, and multi-step workflows using an XState-inspired state machine approach.

• - TypeScript/Node.js backend agents and services
• - Full-stack AI applications with React frontends

The foundational framework for building LLM-powered applications with a massive integration ecosystem. LangChain provides composable abstractions for prompt templates, output parsers, chains, retrievers, and tool integration. While LangGraph is now the recommended path for agents, LangChain remains the backbone for model integrations, tool definitions, and the broader ecosystem.

• - LLM-powered applications with rich integrations
• - Chain-based workflows combining multiple LLM calls

TypeScript-first SDK for building AI-powered web applications with streaming, tool use, structured outputs, and multi-model support. The Vercel AI SDK provides three layers: AI SDK Core for server-side LLM calls, AI SDK UI for React/Svelte/Vue chat hooks, and AI SDK RSC for React Server Components streaming. It is the most popular way to integrate AI into web applications.

• - AI-powered web applications with streaming UIs
• - Chat interfaces with tool use and multi-turn conversations

React component library for building AI copilot experiences with in-app chat, context awareness, and agent integration. CopilotKit provides ready-made React components and hooks that let you embed AI assistants directly into your application, with the ability to read app state, take actions, and interact with LangGraph or other agent backends.

• - In-app AI copilots and assistants
• - React-based AI chat interfaces

Open-source platform for building AI workflows visually with a drag-and-drop interface. Dify combines a visual workflow builder, RAG pipeline, model management, and observability into a single platform. It supports chatbots, agents, text generation, and complex workflow applications, with support for 100+ model providers.

• - No-code AI workflow building for business teams
• - RAG applications with visual pipeline configuration

Open-source drag-and-drop tool for building LLM flows and AI agents with a visual node-based interface. Built on top of LangChain.js, Flowise makes it easy to prototype and deploy chatbots, RAG systems, and agent workflows without writing code. Each node represents a LangChain component that can be configured and connected visually.

• - Visual LLM workflow prototyping and building
• - Chatbot creation without coding

Desktop IDE for visually building, debugging, and testing AI agent graphs with a node-based editor. Rivet was built by Ironclad for their own AI needs and open-sourced as a tool for teams to collaborate on AI agent development. It focuses on the development experience: visual debugging, real-time execution tracing, and team-friendly project management.

• - Visual agent prototyping and design
• - Debugging complex AI agent logic step by step

Open protocol that standardizes how LLM applications connect to external data sources, tools, and services. MCP defines a client-server architecture where MCP servers expose tools, resources, and prompts through a standard interface, and MCP clients (like Claude, IDEs, and agent frameworks) can discover and use them. Think of it as a 'USB-C for AI' — one standard connector for all integrations.

• - Standardized tool integration for LLM applications
• - Connecting AI assistants to databases, APIs, and services

Google's open protocol for agent-to-agent communication and interoperability across different frameworks and organizations. A2A enables agents built with different frameworks to discover each other's capabilities, negotiate interaction modes, and collaborate on tasks. It complements MCP (which connects agents to tools) by focusing on agent-to-agent coordination.

• - Cross-framework agent communication and collaboration
• - Agent discovery and capability advertisement