15 · Agent Development

Agent Frameworks

Agent frameworks give developers reusable building blocks for creating AI systems that can use models, tools, state, memory, and multi-step workflows.

Main mental model

MODEL+INSTRUCTIONS+TOOLS+STATE / MEMORY+AGENT LOGIC↓AGENT FRAMEWORK↓AI APPLICATION

Core: A framework helps build the system. Framework is not the model.

Why frameworks exist

WITHOUT A FRAMEWORK

Developers manually manage:

tool calls
state and agent loops
routing and retries
multi-agent logic

→

WITH A FRAMEWORK

Reusable orchestration components make common system patterns easier to assemble and maintain.

Simple systems may not need a framework.

Major framework landscape

LangGraph

LangChain

Style: graph and state orchestration

Typical fit: controllable, long-running agent workflows

LangChain

Style: general agent tooling

Typical fit: composing model, tool, and retrieval components

OpenAI Agents SDK

OpenAI

Style: provider-native SDK

Typical fit: code-first agents with handoffs and tools

Google ADK

Google

Style: provider-native SDK

Typical fit: agents built around Google’s model ecosystem

AutoGen

Microsoft

Style: multi-agent orchestration

Typical fit: agent-to-agent conversations and collaboration

CrewAI

Style: role-based multi-agent

Typical fit: teams of specialized agents with assigned roles

Semantic Kernel

Microsoft

Style: enterprise agent tooling

Typical fit: integrating AI capabilities into existing applications

Framework categories

GRAPH / STATE ORCHESTRATION

Explicit steps, transitions, and state.

SDK / PROVIDER-NATIVE

Tools and patterns close to a model provider.

MULTI-AGENT

Patterns for several agents to coordinate.

GENERAL AGENT TOOLING

Reusable pieces for building agent applications.

Categories overlap; a framework can fit more than one.

What a framework is not

Framework vs Model

Model: provides AI reasoning and generation capability.

Framework: organizes how models, tools, state, and workflows work together.

Claude / GPT / Gemini models
LangGraph / Agents SDK framework / SDK

Framework vs Agent Product

Framework: used to build AI applications.

Agent product: a ready-to-use experience for users.

LangGraph ≠ Codex
AutoGen ≠ Manus

Framework vs Workflow Platform

Framework: a code-first development layer.

Workflow platform: often visual or low-code orchestration.

LangGraph vs n8n / Dify

There can be overlap between these categories.

How to think about choosing

Need code-level control→ look at a framework or SDK

Need multi-agent patterns→ look at multi-agent-oriented options

Need provider-native integration→ look at a provider SDK

Need visual workflow→ see Module 16

Quick comparison

Framework	Style	Code-first	Multi-agent support	Provider relationship	Typical fit
LangGraph	Graph / state	Yes	Possible	Multi-provider	Controlled workflows
LangChain	General tooling	Yes	Possible	Multi-provider	Composable agent apps
OpenAI Agents SDK	Provider SDK	Yes	Handoffs	OpenAI-native	OpenAI agent systems
Google ADK	Provider SDK	Yes	Supported	Google-native	Google model ecosystem
AutoGen	Multi-agent	Yes	Core focus	Multi-provider	Agent collaboration
CrewAI	Role-based	Yes	Core focus	Multi-provider	Specialist agent teams
Semantic Kernel	Enterprise tooling	Yes	Supported	Multi-provider	Existing applications

Remember this

An agent framework helps developers organize models, tools, state, and workflows into an AI agent system.

If you can explain framework vs model, framework vs agent product, framework vs workflow platform, and why frameworks exist, you have the core idea.