SDKs and libraries for building AI agents. Compare CrewAI, LangGraph, pydantic-ai, Mastra, and the newer entrants shaping how multi-agent, graph, and workflow-based agents get built.
10 frameworks
Role-playing multi-agent framework in Python. Define agents with roles, goals, and tools; assemble them into crews that collaborate on tasks. Widely adopted as the default "let me try multi-agent" starting point.
Graph-based agent framework from LangChain. Build stateful multi-actor applications as directed graphs where nodes are agents or tools and edges encode control flow. The default for anything that needs branching or human-in-the-loop.
OpenAI's official Python SDK for agents. Ships handoffs, guardrails, tracing, and a ReAct-style loop built around the Responses API. The reference for anyone starting on OpenAI models.
Type-safe agent framework from the Pydantic team. Structured outputs, dependency injection, and model-agnostic tool calls with Python type hints as the contract. First-class TypedDict validation of model output.
TypeScript-first agent framework. Agents, workflows, tools, and evals in one library, aimed at devs building agents inside Next.js/Node apps rather than Python notebooks.
TypeScript framework for durable AI workflows and agents. Long-running tasks with automatic retries, checkpointing, and observability. Boundary case: also runs as a hosted platform, but the code is the entry point.
Build and evaluate agents with MCP, ACP, and A2A protocol support. Emphasis on the protocol layer so agents can talk to each other and to tool servers without bespoke plumbing.
Agent-native platform for stateful, durable agents on WebAssembly. Rust core with SDKs in multiple languages. Optimized for agents that need to survive process restarts and cluster failovers.
Microsoft's multi-agent conversation framework. Agents talk to each other in turns, human-in-the-loop optional. AutoGen Studio ships a visual builder on top. Widely used in research and enterprise pilots.
Python framework for building agents with memory, knowledge, and tools. Ships an agent playground UI and API server. Marketed as "5,000× faster than LangGraph" for hot-path agent instantiation.