AI Agent SDK: OpenAI Agents SDK vs Claude Agent SDK vs Google ADK (2026)
Which AI agent SDK to build on in November 2026: the OpenAI Agents SDK, the Claude Agent SDK, Google's Agent Development Kit, Microsoft Agent Framework, Pydantic AI and AWS Strands compared by language, primitives and hosting, plus when SPIRITT ships the agent for you.
An AI agent SDK is a code library that gives you the agent loop, tool calling, handoffs, memory and tracing, so you can build an AI agent in your own application. In November 2026 the main ones are the OpenAI Agents SDK, the Claude Agent SDK, Google's Agent Development Kit (ADK), Microsoft Agent Framework, Pydantic AI and AWS Strands Agents, all free and open source, and you pay for the models and the hosting. SPIRITT, the operating system for AI transformation, is the option when you want the finished agent instead of the library: it codes the agent, tests it, deploys it on its own cloud computer, connects it to 1,500+ apps and runs it 24/7 with budgets and approvals. This guide compares the SDKs by language, core primitives, best fit and where each one runs, using each project's official documentation.
AI agent SDKs at a glance (checked November 2026)
| SDK | Languages | Core primitives | Best for | Price |
|---|---|---|---|---|
| OpenAI Agents SDK | Python, TypeScript | Agents, handoffs and agents as tools, guardrails, sessions, built-in tracing, sandbox agents, realtime voice agents | Lightweight multi-agent apps, voice agents, teams on OpenAI | Free, open source; pay for model usage |
| Claude Agent SDK | Python, TypeScript | Claude Code's agent loop as a library: built-in file, command and web tools, permissions, hooks, subagents, sessions | Agents that read and edit files, run commands and code | Free SDK; pay for Claude API usage |
| Google ADK | Python, TypeScript, Go, Java, Kotlin | Agents, graph workflows (ADK 2.0), sequential, parallel and loop workflows, session state, agent routing | Multi-agent systems on Gemini and Google Cloud | Free, open source; pay for models and cloud |
| Microsoft Agent Framework | Python, .NET | Agents and multi-agent workflows; successor to AutoGen and Semantic Kernel | Azure and .NET teams | Free, open source; pay for models and Azure |
| Pydantic AI | Python | Typed agents, structured outputs, dependency injection, many model providers | Python teams who want type safety | Free, open source |
| AWS Strands Agents | Python, TypeScript | Model-driven agent loop, swarm and graph patterns; pairs with Bedrock AgentCore | AWS teams | Free SDK; AgentCore runtime from $0.0895 per vCPU-hour |
| SPIRITT | Plain English (writes code in any language when needed) | Agent, tools, 1,500+ integrations, cloud computer, hosting, schedules, approvals, budgets | Teams that want a running agent, not a library | Free to start; Starter $32, Pro $256, Max $1,024 a month |
The OpenAI Agents SDK
The OpenAI Agents SDK describes itself as a production-ready upgrade of OpenAI's earlier Swarm experiment, with a deliberately small set of primitives: agents (models with instructions and tools), agents as tools and handoffs (one agent delegates to another), and guardrails (input and output validation that runs in parallel and fails fast). It ships with built-in tracing to visualize and debug runs, sessions for memory, sandbox agents that run inside isolated workspaces, and realtime voice agents.
Strengths: quick to learn, Python-first, very good for voice and handoff-heavy apps. It also works with other model providers through adapters such as LiteLLM. Watch out for: you host, scale, monitor and secure the agent yourself.
The Claude Agent SDK
Anthropic's Claude Agent SDK gives you the same tools, agent loop and context management that power Claude Code, programmable in Python and TypeScript. Out of the box the agent can read, write and edit files, run commands and search the web, with permissions, hooks at key points in the lifecycle, subagents for focused subtasks and resumable sessions. Anthropic also offers Managed Agents, where it hosts the agent loop in a cloud sandbox.
Strengths: the most capable starting point for agents that work with files, code and the command line. Watch out for: built around Claude models; the agent runs in a process you operate unless you choose Managed Agents.
Google Agent Development Kit (ADK)
Google's ADK supports Python, TypeScript, Go, Java and Kotlin. Version 2.0 added graph workflows with routes, retries, timeouts and human input, alongside sequential, parallel and loop workflow agents and agent routing. Agents share context through session state and deploy to Google Cloud.
Strengths: the widest language support and structured multi-agent workflows. Watch out for: you get the most from it inside Google Cloud and with Gemini models. Our guides to Google's AI agent and Vertex AI Agent Builder cover the hosted side.
Microsoft Agent Framework, Pydantic AI and Strands
- Microsoft Agent Framework brings AutoGen's multi-agent ideas and Semantic Kernel's enterprise features into one framework for Python and .NET, with a natural path to Azure. See Azure AI agents.
- Pydantic AI is the choice for Python developers who want typed, validated outputs and a clean way to swap model providers.
- AWS Strands Agents keeps the loop model-driven and pairs with Bedrock AgentCore for runtime, memory, browser and code interpreter. See AWS AI agents.
SDK or platform: what an SDK leaves to you
An SDK gives you the loop. A production agent needs much more around it:
- Hosting and uptime. A server or container that runs 24/7, restarts on failure and scales.
- Integrations and credentials. OAuth for every app the agent touches, secrets storage and key rotation.
- A safe place to act. A sandbox or browser for computer use, isolated from your other systems.
- Observability and evals. Traces, cost per run and regression tests before each change. See our guide to AI agent evaluation.
- Guardrails people can manage. Approvals for risky actions, budgets and an audit trail that a non-developer can read.
This is the work SPIRITT does for you. You describe the agent in plain English; SPIRITT writes the code, composes models from several labs within each task, connects 1,500+ apps, deploys the agent on its own isolated cloud computer and runs it on a schedule or on events. Every workspace is isolated, keys are injected at the network boundary so agents never see them, and each integration has allow, ask and never rules. You can still bring code: SPIRITT runs other agents such as Claude Code and Codex in the same place, and can host an agent you built with any of these SDKs.
How to choose an AI agent SDK
- Your model and cloud. OpenAI shop: OpenAI Agents SDK. Claude and code-heavy agents: Claude Agent SDK. Google Cloud: ADK. Azure and .NET: Microsoft Agent Framework. AWS: Strands.
- Your language. Only ADK covers Go, Java and Kotlin; Microsoft Agent Framework covers .NET; most others are Python and TypeScript.
- Single agent or many. Handoffs (OpenAI), subagents (Claude), graph workflows (ADK) and swarm patterns (Strands) suit different team shapes. Our guide to AI agent design patterns explains when each pattern fits.
- Who maintains it. If no engineer will own hosting, upgrades and on-call, a platform that runs the agent for you is cheaper in practice.
FAQ
What is the best AI agent SDK?
There is no single best one. The OpenAI Agents SDK is the lightest for handoffs and voice, the Claude Agent SDK is strongest for agents that work with files and code, and Google ADK has the widest language support. For a running agent without maintaining code, SPIRITT builds and hosts it for you.
What is the OpenAI Agents SDK?
It is OpenAI's open-source library for building agentic apps with a few primitives: agents, handoffs and agents as tools, and guardrails, plus sessions, tracing, sandbox agents and realtime voice agents. It is free; you pay for model usage.
What is the Claude Agent SDK?
It is Anthropic's library that exposes Claude Code's agent loop, built-in tools, permissions, hooks and subagents to your own Python or TypeScript application. Install it with pip as claude-agent-sdk or with npm as @anthropic-ai/claude-agent-sdk.
Is the OpenAI Agents SDK free?
Yes. The SDK is open source and free; you pay OpenAI or another model provider for the tokens your agent uses, plus your own hosting.
Can I use the OpenAI Agents SDK with other models?
Yes. It supports other providers through adapters such as LiteLLM and Any-LLM, though some features are built around OpenAI's Responses API.
What is the difference between an AI agent SDK and an agent framework?
The terms overlap. An SDK usually means a vendor's library with a small set of primitives, while a framework such as LangGraph or CrewAI adds more structure for graphs, roles and state. Both leave hosting, integrations and monitoring to you.
Do I need an SDK to build an AI agent?
No. Developers use SDKs to embed agents in their own products. To automate your own business work, you can describe the agent to SPIRITT in plain English; it writes the code, deploys and runs it.
How we checked
SPIRITT publishes this guide. SDK facts come from the official documentation for the OpenAI Agents SDK, the Claude Agent SDK, Google ADK, Microsoft Agent Framework, Pydantic AI and AWS Strands Agents, checked on November 22, 2026. AWS AgentCore pricing comes from AWS's pricing page in November 2026. SPIRITT facts come from spiritt.ai and the SPIRITT product.