AI agent frameworks compared: Claude Agent SDK, OpenAI Agents SDK, Google ADK and more

AI agent frameworks are libraries that run the agent loop for you: send a task to a model, execute the tools it asks for, return the results and repeat until the job is done. Options in October 2026 include the vendor SDKs (Claude Agent SDK from Anthropic, OpenAI Agents SDK, Google's Agent Development Kit) and independent or platform frameworks such as LangGraph, CrewAI and Microsoft Agent Framework.

They differ most in three ways: which models they run, what comes built in (some ship file and shell tools, others give you primitives only), and how they orchestrate multiple agents. This guide compares them from their official docs, without ranking them, so you can match one to your project.

What does an AI agent framework do?

Anthropic's Building effective agents post (December 2024) separates two kinds of systems:

• "Workflows are systems where LLMs and tools are orchestrated through predefined code paths."
• "Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage."

Underneath both is tool calling: the model returns a structured request to run a function, your code runs it, and the result goes back to the model. A framework handles that loop, plus the parts around it that get tedious to write by hand:

• Tool definitions: turning functions into schemas the model can call.
• State: sessions or memory so a conversation can continue or resume.
• Orchestration: handing work between agents, or running steps in sequence, in parallel or in loops.
• Safety: guardrails, permissions and human approval steps.
• Observability: tracing each model call and tool call so you can debug a run.

Most frameworks also connect to MCP servers, so tools written once can be shared across agents.

Do you need a framework at all?

Often not at first. Anthropic's post says: "We suggest that developers start by using LLM APIs directly: many patterns can be implemented in a few lines of code." It warns that frameworks "often create extra layers of abstraction that can obscure the underlying prompts and responses," and adds: "If you do use a framework, ensure you understand the underlying code."

Microsoft's Agent Framework overview makes a similar point about agents in general: "If you can write a function to handle the task, do that instead of using an AI agent."

A framework is worth it when you need what it ships: built-in tools, durable sessions, tracing, approvals or multi-agent orchestration you would otherwise build yourself.

How do the main AI agent frameworks compare?

All details below come from each project's official docs or repository, read in October 2026:

What is the Claude Agent SDK?

The Agent SDK overview says it "gives you the same tools, agent loop, and context management that power Claude Code, programmable in Python and TypeScript." The packages are claude-agent-sdk for Python and @anthropic-ai/claude-agent-sdk for TypeScript, and both bundle a native Claude Code binary.

Its distinguishing feature is what comes built in. The agent can read, write and edit files, run commands and search the web without you implementing those tools. Permissions control which tools run automatically, hooks run your code at points in the loop, subagents handle focused subtasks, and skills and memory load from .claude/ folders the same way they do in Claude Code. A minimal TypeScript agent, adapted from the quickstart:

import { query } from "@anthropic-ai/claude-agent-sdk";

for await (const message of query({
  prompt: "Review utils.py for bugs that would cause crashes. Fix any issues you find.",
  options: { allowedTools: ["Read", "Edit", "Glob"], permissionMode: "acceptEdits" },
})) {
  if (message.type === "result") console.log(`Done: ${message.subtype}`);
}

It authenticates with an API key or through Amazon Bedrock, Claude Platform on AWS, Google Cloud's Agent Platform or Microsoft Foundry. The docs note that third-party developers may not offer claude.ai login or rate limits in their products unless previously approved. Our Claude Agent SDK guide covers setup in depth.

What is the OpenAI Agents SDK?

OpenAI's Agents SDK docs describe it as a way to "build agentic AI apps in a lightweight, easy-to-use package with very few abstractions," and as "a production-ready upgrade of our previous experimentation for agents, Swarm." The core primitives are agents (models with instructions and tools), agents as tools or handoffs for delegation, and guardrails for validating inputs and outputs. Sessions, human-in-the-loop steps and built-in tracing sit around them.

The Python package is openai-agents; the TypeScript version installs with npm install @openai/agents zod and adds realtime voice agents. The docs' hello world:

from agents import Agent, Runner

agent = Agent(name="Assistant", instructions="You are a helpful assistant")

result = Runner.run_sync(agent, "Write a haiku about recursion in programming.")
print(result.final_output)

It uses the Responses API by default for OpenAI models. The models page documents LiteLLM and Any-LLM adapters for other providers and says that support is "on a best-effort, beta basis."

What is Google's Agent Development Kit?

The ADK site calls it "the open-source agent development framework that lets you build, debug, and deploy reliable AI agents at enterprise scale." It is available in Python (pip install google-adk), TypeScript (npm install @google/adk), Go, Java and Kotlin, and its TypeScript 2.0 release added graph workflows.

ADK leans toward structured orchestration. Besides model-driven agents, it ships template workflows (sequential, loop and parallel) and graph-based workflows, along with evaluation tooling and a web interface for debugging. Tools can be plain functions, MCP tools or OpenAPI tools. For deployment, the site names "Agent Runtime (Agent Platform), Cloud Run, or GKE" on Google Cloud, and you can also run it on your own infrastructure.

If you build ADK agents with a coding assistant, see the Google agents-cli ADK code skill.

What about LangGraph, CrewAI and Microsoft Agent Framework?

LangGraph. The LangGraph docs describe "a low-level orchestration framework and runtime for building, managing, and deploying long-running" stateful agents. You model your agent as a graph of nodes and shared state, and it focuses on durable execution, streaming, human-in-the-loop steps and memory. The docs say "you don't need to use LangChain to use LangGraph," and point you to LangChain's prebuilt agents if you want a higher-level abstraction.

CrewAI. The CrewAI docs organize work into Crews, "teams of autonomous agents that collaborate to solve specific tasks," and Flows, "structured, event-driven workflows that manage state and control execution." Its README calls it "a standalone Python framework with its own primitives for agents, tasks, crews, flows, tools, and orchestration." It installs with uv tool install crewai.

Microsoft Agent Framework. Microsoft calls it "the direct successor" to Semantic Kernel and AutoGen, "created by the same teams." It combines single agents, an opinionated harness agent for long multi-step tasks, and graph-based workflows, with model clients for Microsoft Foundry, Anthropic, Azure OpenAI, OpenAI, Ollama and others. The Go version is in public preview.

How do you choose an AI agent framework?

No framework fits every project. These questions narrow it down:

• Are you committed to one model vendor? The vendor SDKs track their own platform's features first. The Claude Agent SDK is built for Claude models; the OpenAI and Google SDKs support other models through adapters; LangGraph, CrewAI and Microsoft Agent Framework are built to mix providers.
• Does the agent need to work on files and run commands? The Claude Agent SDK ships those tools and a permission system. With the others, you define such tools yourself or connect them through MCP.
• How much control over the flow do you want? Handoff-style SDKs let the model decide the path. Graph and workflow frameworks (LangGraph, ADK workflows, Microsoft workflows, CrewAI Flows) let you fix the path in code. Anthropic's workflow versus agent distinction maps onto this choice.
• Which language does your team use? Python is covered by all six; TypeScript by Claude, OpenAI, Google and LangGraph; Go, Java, Kotlin and .NET by ADK or Microsoft.
• Where will it run? ADK has one-command deploys to Google Cloud, Microsoft's framework fits Azure and Foundry, and the Claude SDK runs wherever you can run its bundled binary.

Prototype the same small task in two candidates before committing. The differences show up quickly in how tools, state and errors feel.

What does every framework still leave to you?

A framework runs the loop; it does not make the agent correct or safe. You still own:

• Prompts and instructions. Every framework starts from a system prompt you write and refine.
• Evaluation. You need test tasks with known outcomes to know whether a change helped.
• Permissions and blast radius. Decide which tools may run without approval, especially anything that writes, deletes or spends money.
• Untrusted input. Tool results from web pages, emails or issues can carry instructions aimed at the model. Our guide to prompt injection in coding agents covers defenses that apply to any framework.

Where should your agent's prompts and skills live?

Agent instructions and skills change often and are easy to lose across repos. Kipt keeps prompts and Agent Skills as Markdown in a private GitHub repo you own, and installs every skill into Claude Code or Codex with one command. The Claude Agent SDK loads skills from the same .claude/ folders that Claude Code uses.

What should you read next?

Start with what is tool calling to understand the loop every framework wraps, then what is MCP to share tools between agents. If you choose Anthropic's SDK, the Claude Agent SDK guide walks through a full build, and what is an AI coding agent shows the same architecture in tools you may already use.