~55 min read · updated 2026-05-10

Agent frameworks

LangChain, LangGraph, Pydantic AI, Claude Agent SDK, OpenAI Agents SDK, CrewAI, AutoGen — what each is good at, what to pick when.

You can hand-roll the agent loop from module 01 forever. At some point, the runtime features — streaming, parallel calls, retries, observability hooks, structured outputs, multi-agent orchestration — are enough work that picking a framework saves time. This module walks through the major frameworks side-by-side.

This module is being expanded with side-by-side code for the same agent in LangGraph and the Claude Agent SDK (the two most-recommended 2026 starting points).

Coming in the next revision:

  • The landscape: LangChain, LangGraph, LlamaIndex, CrewAI, AutoGen, Pydantic AI, Claude Agent SDK, OpenAI Agents SDK, Vercel AI SDK, Mastra, Microsoft Semantic Kernel.
  • The decision tree — when to pick each.
  • A side-by-side: the same agent written in LangGraph (graph-based state machine) vs Claude Agent SDK (linear loop + MCP tools).
  • Migration paths between frameworks.
  • The pattern: frameworks for orchestration, MCP for tools, model-vendor SDK for vendor-specific features. Three layers, three responsibilities.

Next: Module 07 — Multi-agent systems.