Official OpenAI & Claude Agent SDKs: How to Build Directly Without Complex Frameworks

Until now, building your own AI agent meant learning complex and heavy external frameworks first. But the development landscape is changing completely, as AI companies like OpenAI, Anthropic, and Google have started releasing their own official SDKs.

You no longer need to waste time studying complicated configurations. By simply importing the lightweight official libraries provided by each company, agent behavior and tool integration are handled seamlessly.

In this post, I will walk through the recent trend of official SDKs from big tech companies and break down the key features in an easy-to-understand way, even for beginners.

1. OpenAI Agents SDK: A Lightweight and Powerful Multi-Agent Tool Beyond Swarm

Following the experimental project Swarm, which garnered much attention, OpenAI has finally released their official multi-agent development library, openai-agents! Without the need to load heavy third-party frameworks, you can now build AI assistants that run cleanly with just a few lines of Python code.

This SDK is intuitive enough to start with, requiring knowledge of only three core concepts instead of complex abstractions.

Check out this light code to see how an actual agent is structured.

# PyPI openai-agents 패키지 기준
from openai_agents import Agent, Runner, SQLiteSession

# 에이전트가 활용할 도구 정의
def get_current_time():
    return "현재 시간은 오후 4시 25분입니다."

# 에이전트 선언
assistant = Agent(
    instructions="시간 정보를 알려주는 비서입니다.",
    tools=[get_current_time]
)

# 세션 선언 및 에이전트 실행
session = SQLiteSession(database_path="chat_history.db")
response = Runner.run_sync(
    agent=assistant,
    prompt="지금 몇 시야?",
    session=session
)
print(response.text)

The safety features added in the latest update are also very interesting. By using SandboxAgent and SandboxRunConfig, the agent can test file editing or shell commands within an isolated Unix environment. You can confidently delegate automated tasks, knowing you can verify if the code works correctly inside a safe sandbox without worrying about damaging your computer's OS environment.

2. Claude Agent SDK: Combining Computer Control with the SKILL.md Standard

Anthropic's Claude Agent SDK is a tool that gives AI control over a computer. It is the official library that allows developers to code the core operating principles of Claude Code, the terminal AI tool that has recently gained significant attention among developers.

Using this SDK, you can use powerful local environment control tools such as reading and writing files, system searches, and executing Bash commands by default. It supports both JavaScript and Python environments, and you can get started immediately by installing the @anthropic-ai/claude-agent-sdk and claude-agent-sdk packages, respectively.

A key feature to note is the native support for the open standard agent skill specification, or SKILL.md. This specification is a technique for bundling guidelines or tool permissions that the AI uses into a standalone Markdown file. The AI keeps only the summary information of the necessary skills on hand, and only reads the full guide when that task is actually required, significantly reducing wasted tokens.

SKILL.md files can be structured intuitively, with setting metadata at the top and detailed instructions at the bottom, as shown in the example below.

---
name: deploy_project
description: "프로젝트를 배포 서버에 업로드하고 동작 상태를 검증하는 스킬"
allowed-tools: [bash, write_file]
---

# 프로젝트 배포 가이드
이 스킬이 활성화되면 다음 순서대로 실행하세요.
1. `npm run build` 명령어로 빌드가 성공하는지 검증합니다.
2. 배포 셸 스크립트를 작동시켜 빌드 폴더를 전송합니다.

There's no need to force-feed all prompts to the agent, making it heavy. By simply placing a guidebook dedicated to specific tasks in a folder, you can make Claude work smartly, as if it’s pulling a guide from a drawer whenever it needs it.

3. Google ADK 2.0: Visual Debugging and Easy Deployment in One

Google's Agent Development Kit, google-adk 2.0, focuses above all on building a stable, production-ready environment. The design stands out for cleanly organizing the often-tangled execution flow of agents. It is particularly scalable, allowing you to freely combine and use various models without being tied to a specific AI model.

The key features of Google ADK 2.0 include:

Its most attractive strength is that it drastically solves the perennial developer headaches of debugging and deployment. With just a short command, you can run a local web server to visually track the agent's thought process and tool execution history in detail.

Everything flows seamlessly from testing to cloud deployment with a single command.

# 로컬에서 에이전트 동작 시각화하기 (localhost:8000)
adk web

# 준비가 완료되면 구글 클라우드에 즉시 배포하기
adk deploy

Once you verify that it runs without errors in your local environment, you can deploy it to Google Cloud and start the service with just a single command, without any complex configuration.

Which SDK should you start with?

Choose the official SDK that best fits the nature of the agent you are building.

There's no need to struggle with learning heavy and complex third-party frameworks. Pick the official SDK that’s just right for you, run a simple script, and enjoy the fun of building agents!