# Agent frameworks at a glance: which one for which job

A one-page decision guide to agent frameworks. The full comparison, with code for each, is in the agent frameworks article.

## Overview

Agent frameworks differ less in what they can do than in how much they decide for you. Libraries give you building blocks, runtimes run the loop with saved state, harnesses add planning and sub-agents, and platforms add tracing, evaluation and hosting.

## Decide by the shape of the task

| Your task | Good fit | Why |
|---|---|---|
| One model, a few tools | LangChain `create_agent`, [OpenAI Agents SDK](/tools/openai-agents-sdk), [PydanticAI](/tools/pydantic-ai) | Small loop, little framework code |
| A process with approvals and resume after failure | [LangGraph](/tools/langgraph), [Microsoft Agent Framework](/tools/microsoft-agent-framework) workflows | Explicit graph and checkpoints |
| Long research or coding tasks | Deep Agents, [Claude Agent SDK](/tools/claude-agent-sdk) | Planning, files and sub-agents built in |
| Role-based business processes | [CrewAI](/tools/crewai) | Roles, tasks and crews read like a team chart |
| Retrieval-heavy agents | [LlamaIndex](/tools/llamaindex) | Strongest data and RAG tooling |
| Fewer steps through code actions | [smolagents](/tools/smolagents) | The model writes Python that calls several tools |
| Measurable tasks to optimise | [DSPy](/tools/dspy) | Optimises prompts against a metric |
| Agents across teams or companies | [Google ADK](/tools/google-adk) or any A2A framework | Native A2A |
| TypeScript web stack | [Mastra](/tools/mastra), LangGraph.js | Same language as the app |

## Decide by your platform

Azure and .NET teams usually choose Microsoft Agent Framework, which replaced AutoGen and Semantic Kernel in April 2026. Google Cloud teams choose ADK, AWS teams Strands Agents, and teams that want portability choose LangGraph or a small SDK with tools behind [MCP](/tools/mcp).

## What to decide before the framework

1. **Tracing:** choose [LangSmith](/tools/langsmith), [Langfuse](/tools/langfuse) or an [OpenTelemetry](/tools/opentelemetry) tool on day one.
2. **Evaluation:** a fixed test set and a pass threshold, with tools such as [DeepEval](/tools/deepeval) or [promptfoo](/tools/promptfoo).
3. **Tool boundary:** tools behind MCP so the framework can change later.

## Read the full comparison

The article **[Agentic AI frameworks explained: the Lang family and every major alternative](/articles/agentic-ai-frameworks-langchain-langgraph-and-every-alternative)** explains every framework, the difference between LangChain, LangGraph, LangSmith, Langflow and Langfuse, and shows the same agent in several of them.

## Frequently asked questions

### Which framework is the most popular?

LangChain and LangGraph have the largest ecosystems; CrewAI, the OpenAI Agents SDK, Google ADK and Microsoft Agent Framework each have large communities in their own platforms.

### Can I switch frameworks later?

Yes, if tools sit behind MCP and prompts and evaluations live outside the framework. The agent loop itself is small and quick to rewrite.

## Summary

Pick by task shape first, platform second, and settle tracing, evaluation and the tool boundary before writing agent code.

## Back to the topic

This page is part of **[Agentic AI](/topics/agentic-ai)**, which maps every concept, tool, simulation and article on the subject.
