13 Frameworks and SDKs for Building AI Agents

Building an AI agent today rarely means starting from scratch. Developers can use frameworks and SDKs that already handle much of the infrastructure around the model – tool calling, memory, orchestration, state management, multi-agent workflows, tracing, human approval, and deployment.

SDKs and agent frameworks make the whole process much easier. For example, you want to add GPT to your app. You could manually construct HTTP requests, handle authentication, parse responses, manage errors, and so on. Or you can use an AI SDK and write a few lines of code like client.responses.create(...). So much of the technical plumbing is handled by the SDK, while an agent framework can additionally handle higher-level orchestration, state, and workflows.

As a result, we get the following structure: an AI model provides the intelligence, an API is the interface for communicating with it, an SDK provides convenient tools for using that API, and an agent framework provides higher-level components for building and orchestrating agent behavior.

The options now range from lightweight runtimes to graph-based orchestration systems and specialized frameworks for RAG, multi-agent systems, and realtime voice. So which one should you use? Here are the major options and what each is best suited for.

OpenAI Agents SDK – a lightweight agent runtime for building and running multi-agent workflows with tools, agent-to-agent handoffs, guardrails, memory/session management, tracing, human approval, MCP, sandbox agent, and real time voice support. It is provider-agnostic, supporting OpenAI APIs and 100+ other LLMs.

You can build agents using the Python SDK package and TypeScript version for JavaScript and Node.js applications with it. The Python repo already has ~28.5k stars.

LangGraph is a low-level orchestration framework rather than a high-level SDK, built for long-running, stateful agents. It supports durable execution, persistence, memory, human-in-the-loop, and production deployment. You can use Python and TypeScript for developing agents. LangGraph is one of the more mature options for complex stateful systems, with ~39k GitHub stars.

Google Agent Development Kit (ADK) is Google’s open-source, code-first framework for building, evaluating, and deploying agents. ADK 2.0 includes a graph-based workflow runtime with routing, loops, retries, state management, human-in-the-loop, nested workflows, and structured agent-to-agent delegation. With it, you can develop agents in Python.

Microsoft Agent Framework (MAF) is Microsoft’s open, multi-language framework for building production-grade agents and multi-agent workflows. It is available for Python and .NET. Its features include graph-based orchestration, multiple model providers, A2A, checkpointing, observability, human-in-the-loop, and local or cloud deployment. Microsoft also provides migration guides for moving from AutoGen and Semantic Kernel to MAF.

Pydantic AI is a type-safe agent framework from the creators of Pydantic, with dependency injection, structured outputs, tools, MCP, durable execution, graphs, and observability. A strong option for production Python stacks where typing and validation matter most.

CrewAI is an open-source framework built around crews, agents, tasks, and flows. Crews let role-based agents collaborate and delegate tasks, while Flows add more control through event-driven workflows, state management, and conditional branching. It’s a good fit when you need several agents working together using Python rather than a minimal runtime for a single agent.

Agno – a framework and Python SDK for building complete agent platforms, rather than just individual agents. You build agents with the Agno SDK, run them as services through the AgentOS runtime, and manage everything from the AgentOS UI. It comes with memory, tracing, scheduling, human approval, JWT-based access control, and your own storage for data and sessions.

Strands Agents SDK is available for both Python and TypeScript. It takes a simple, model-driven approach to building agents and works with any model or cloud, although both SDKs use Amazon Bedrock by default. You can switch to Anthropic, OpenAI, Gemini, Ollama, and other providers without changing the overall agent architecture.

Mastra – a TypeScript-native framework for building AI applications and agents. It brings agents, graph-based workflows, memory, RAG, MCP, evals, and observability into one stack, then lets you integrate everything with React, Next.js, or Node.js, or deploy it as a standalone server. It is especially useful for developers already working in the TypeScript ecosystem.

Hugging Face smolagents is a lightweight Python library for building agents with minimal abstraction. Its CodeAgent follows a ReAct-style loop but writes each action as a Python code snippet, calling tools as regular functions and executing the code in a sandbox. For more conventional tool use, there’s also a ToolCallingAgent that generates JSON or text-based tool calls. It works with local models and providers such as OpenAI and Anthropic.

LlamaIndex / LlamaAgents is a framework and toolkit for building agents that work with private data and documents, connecting to sources such as PDFs, APIs, SQL databases, and other files. It structures that data into indices or graphs and retrieves relevant context when a query arrives. Its main strengths are RAG, document processing, data extraction, and end-to-end document agents built with event-driven Workflows. Available for both Python and TypeScript

CAMEL-AI – An open-source Python framework for building and studying multi-agent systems. You create agents with specific roles, models, tools, and memory, then organize them into societies where they communicate and collaborate on tasks. It’s especially strong for role-playing agents, workforce orchestration, large-scale simulations, synthetic data generation, and research into agent behavior.

LiveKit Agents is a specialized framework for building realtime voice and multimodal agents that run as programmable participants on a server. An AgentSession manages the interaction with each user, while you can either connect STT, an LLM, and TTS into a voice pipeline or use a realtime model directly. LiveKit handles WebRTC communication, turn detection, tools, scheduling, telephony, and multi-agent handoffs. You can use it to build agents in Python and JavaScript/TypeScript.

An AI agent framework provides higher-level components for building and orchestrating AI agents. Depending on the framework, these can include tool use, memory, state management, workflows, multi-agent coordination, human-in-the-loop controls, tracing, and deployment. Examples include LangGraph, CrewAI, Pydantic AI, and Microsoft Agent Framework.

An AI agent SDK (Software Development Kit) is a collection of libraries, functions, documentation, and other developer tools that makes it easier to build agents without handling all the underlying API interactions manually. OpenAI Agents SDK and Strands Agents SDK are examples. Some tools combine SDK capabilities with broader agent frameworks or runtimes.

An SDK primarily gives developers convenient tools and abstractions for working with APIs and building applications. An agent framework typically adds higher-level architecture for orchestrating agent behavior, including workflows, state, memory, tools, and multi-agent coordination. The distinction is not always strict: some projects combine both approaches.

There is no single best option. LangGraph is designed for complex stateful workflows, Pydantic AI emphasizes type-safe Python development, CrewAI focuses on role-based multi-agent systems, Mastra is built for TypeScript applications, LlamaIndex is strong for data and RAG agents, and LiveKit Agents specializes in realtime voice and multimodal agents. The best choice depends on the application and development stack.

Several options support both Python and TypeScript, including OpenAI Agents SDK, LangGraph, Strands Agents SDK, LlamaIndex, and LiveKit Agents. Other tools specialize in one ecosystem: Pydantic AI, CrewAI, Agno, smolagents, and CAMEL-AI are primarily Python-focused, while Mastra is TypeScript-native.