AI Frameworks · Comparison

Which AI Framework Should You Use?

⏱ 10 min read🎯 All levels🖼 2 diagrams🧪 1 live demo

There are now many frameworks for LLM apps and agents, and they overlap. The good news: the core ideas (prompts, tools, retrieval, loops, memory, tracing) are the same everywhere. Learn them once with LangChain and LangGraph, and you can pick up any other framework in a day. Here is how they compare, and how to choose.

In this session

The same five building blocks in every framework
Nine popular frameworks side by side
An interactive chooser
When to use no framework at all

01The same building blocks everywhere

Every framework packages these five things; they differ in style and defaults.

Learn concepts, not just APIs. Framework APIs change every few months; the ideas in GenAI Course Modules 8–9 and these notes do not.

02Nine frameworks side by side

Framework	Core idea	Best for	Languages
LangChain	standard parts + `create_agent`	LLM apps, RAG, single agents	Python, JS/TS
LangGraph	graph of nodes, edges and saved state	custom workflows, approvals, multi-agent	Python, JS/TS
LlamaIndex	data framework: parse, index, query	chat with documents, knowledge assistants	Python, TS
CrewAI	role-based “crews” of agents with tasks	quick multi-agent prototypes (research, content)	Python
Microsoft Agent Framework	successor to AutoGen and Semantic Kernel	Azure and .NET shops, multi-agent chat	Python, .NET
OpenAI Agents SDK	agents, handoffs, guardrails, tracing	lightweight agents on OpenAI models	Python, JS/TS
Google ADK	Agent Development Kit, agent hierarchies, A2A	Google Cloud / Gemini, multi-language teams	Python, Java and more
Pydantic AI	type-safe agents with validated outputs	structured extraction, typed Python codebases	Python
Claude Agent SDK	the agent loop behind Claude Code: files, shell, subagents	coding and computer-task agents on Claude	Python, TS

Landscape as of late 2026; check each project’s docs for its current version.

03Choose with your requirements

Demo · What are you building?

Question to ask	Why it matters
Do I need loops or approvals?	if not, a chain or plain SDK call is enough
Which models must I support?	single-vendor SDKs lock you to one provider
How will I debug it?	tracing (LangSmith, provider dashboards, OpenTelemetry) is essential for agents
Who maintains it?	large, active projects get security fixes and new model support quickly
Can I test it?	explicit workflows are much easier to unit-test than free-roaming agents

04When to use no framework

Frameworks save time, but every layer adds something to learn and debug. Use plain provider SDKs when the app is a single call, a short fixed chain, or performance-critical. Many production teams start with the SDK, then adopt LangGraph when they need checkpoints and approvals.

Choose by how much of the framework you will actually use.

05Summary

All frameworks share the same blocks: model, tools, knowledge, control flow, memory + tracing.
LangChain for parts and simple agents, LangGraph for controlled workflows, LlamaIndex for data-heavy RAG.
CrewAI for quick multi-agent prototypes; Microsoft Agent Framework for Azure/.NET; OpenAI Agents SDK, Google ADK and Claude Agent SDK for their ecosystems.
For one simple call, skip the framework.

06Quick quiz

1. Which framework is the successor to AutoGen and Semantic Kernel?

2. Best fit for strict workflows with human approval?

3. Best fit for a document-heavy knowledge assistant?

4. When should you skip frameworks?

5. What stays the same across frameworks?

07References

Created by Ajeetkumar