16 - Agent Frameworks

Frameworks package the agent loop, tool schemas, state, approvals, multi-agent wiring and tracing so you don't rebuild them for every project - and they hide those mechanics, which is why this module comes after you built the loop by hand. You learn what each major framework is built around, how to choose one (or none), and build the same agent in seven of them.

Learning objectives 8-10 hours (notes + lab)

By the end of this module you will be able to:

Explain what an agent framework adds to a raw model API loop and what it hides, and place frameworks on the layer map
Compare LangChain/LangGraph, OpenAI Agents SDK, Claude Agent SDK, Google ADK, CrewAI, Microsoft Agent Framework and PydanticAI by abstraction, state, human-in-the-loop, durability and model neutrality
Implement human approval and resumable runs in each framework's native style
Choose a framework for a project with a written rationale backed by measurements

Prerequisites

Agent Foundations - the loop, tools, evaluation
Agent Patterns & Multi-Agent - the patterns frameworks implement
Python 3.10+; a local OpenAI-compatible model server or a hosted API

Where This Module Fits

flowchart LR
    C["01 Choosing a<br/>framework"] --> G["02 LangChain /<br/>LangGraph"]
    C --> O["03 OpenAI<br/>Agents SDK"]
    C --> CL["04 Claude<br/>Agent SDK"]
    C --> A["05 Google ADK"]
    C --> CR["06 CrewAI"]
    C --> M["07 Microsoft Agent<br/>Framework"]
    C --> P["08 PydanticAI"]
    G & O & CL & A & CR & M & P -.-> LAB["🧪 Lab: one agent,<br/>many frameworks"]

    style C fill:#e8e2d9,stroke:#ccc4b8
    style LAB fill:#dde4dc,stroke:#b0c4b0

Chapters 2-8 are independent: read chapter 1, then the frameworks you are considering. Running agents in production - durable execution, security, observability, cost - is Production Agents.

Chapter Map

#	Chapter	You will learn	Time
1	Choosing an Agent Framework	What frameworks add and hide, the layer map, comparison, decision guide, when to use none	45 min
2	LangChain and LangGraph	`create_agent` and middleware; state graphs, checkpoints, `interrupt()`/`Command`, Store	50 min
3	OpenAI Agents SDK	Agents, handoffs, guardrails, sessions, approvals with `RunState`, tracing	40 min
4	Claude Agent SDK	The Claude Code harness as a library: built-in tools, permissions, hooks, subagents, skills	40 min
5	Google ADK	`LlmAgent`, Runner and sessions, graph Workflows, callbacks, confirmation, deploy	45 min
6	CrewAI	Crews of role-playing agents, processes, Flows with typed state	40 min
7	Microsoft Agent Framework	Agents, graph workflows with checkpoints, orchestrations, migration from AutoGen/Semantic Kernel	40 min
8	PydanticAI	Typed agents, dependency injection, deferred tools, testing, durable execution	35 min
9	Q&A Review Bank	31 questions across the module	45 min

Code Lab

Lab	What you build	Runs on
One Agent, Many Frameworks	The Lab 13 shop agent with a refund approval step in LangChain, OpenAI Agents SDK, ADK, CrewAI, MAF and PydanticAI (graded on state), plus a Claude Agent SDK version	A local OpenAI-compatible model (laptop) or any hosted API; Claude Agent SDK needs an Anthropic key

Mini-Project

Choose a framework for a real agent at your work, with evidence:

A task set of 20+ inputs with state-based or rubric checks.
Two candidate frameworks, the tools shared as plain functions, one approval step in each.
Task success (with confidence intervals), tokens, latency and lines of framework-specific code for each.
A one-page decision covering deployment, durability, observability, lock-in and upgrade risk - not just the numbers.

Review

Q&A Review Bank - consolidated questions for this module
Module quiz - every Check Yourself question in this module, in course order

Previous: 15 - Agent Patterns & Multi-Agent · Next: 17 - Production Agents

Section Appendix

Summary & Key Terms - a quick recap of this section and its essential vocabulary.