“Claude agents” is not a single product but an umbrella term for five ways in which Anthropic lets Claude work as an agent, meaning with its own steps and tools rather than just an answer. For your own work there is Claude Code with its subagents. For developers there is the Claude Agent SDK, which brings Claude Code into your own Python or TypeScript application as a library, and Claude Managed Agents, where Anthropic runs the agent loop together with an isolated workspace in the cloud; as of October 26, 2026, Managed Agents is in beta. If you want to control everything yourself, you build the loop directly on the Messages API. And for end users there has been Claude Duet since October 21, 2026, a ready-made agent in its own app, so far only by waitlist in the US. Which route fits comes down to one question: do you want to work with an agent yourself, or should an agent become part of your own software?
I run my team of fourteen AI employees in Claude Code, so on the first route. How to build your own agent with Claude without programming is covered in Build an AI Agent With Claude: 3 Ways. How subagents, skills and a whole AI employee fit together in Claude Code is in Claude Subagents, Skills, Agents. This article maps out Anthropic's developer offerings, which neither of those covers.
What Anthropic means by an agent
In its foundational text on agents, Anthropic separates two things. Workflows are systems in which a language model and tools run through predefined code paths. Agents are systems in which the model directs its own steps and tool use. The Agent SDK puts it even more simply: an agent is an application that completes a task by planning its own steps and calling tools that read files, run commands or edit code.
The same text contains the advice I consider the most important: find the simplest solution possible and only increase complexity when needed. Many tasks that sound like “agent” are better served by a fixed workflow. What sets an agent apart from a chatbot or a workflow is covered in more depth in What Is an AI Agent? Explained Simply.
Claude agents at a glance: five offerings
Source: Anthropic's documentation on the Agent SDK and Managed Agents and the Duet pages, as of October 26, 2026. Alongside these, the Claude apps have Cowork, which since September 2026 has been gradually merging into the normal chat; that is the agent for office work without a terminal, described in Claude Cowork: What It Does, Who It's For.
Subagents in Claude Code: agents for your own work
For most business owners this is the right starting point. Claude Code is itself an agent: it plans, reads files, runs commands and checks the result. Subagents are specialized helpers inside it, each with its own context window, its own system prompt, its own tool permissions and, if you like, its own model. Important for cost: according to the documentation, subagents send their own requests, which count toward the same usage limits as your main conversation. According to the documentation, fork mode is also on by default: Claude can start forks on its own that inherit the conversation so far. Your own subagents still start with a fresh context.
In my setup, thirteen of the fourteen employees are defined as subagents in .claude/agents/; Kai, my team lead, runs directly through Claude Code without a definition file of his own. How such a file is built and when a skill is enough instead of a subagent I will not repeat here; that is in Claude Subagents, Skills, Agents, and the finer points for daily operation in Claude Code Subagents: Best Practices.
Claude Agent SDK: Claude Code as a library
According to Anthropic, the Agent SDK gives you the same tools, the same agent loop and the same context management that power Claude Code, programmable in Python and TypeScript. Technically, the SDK runs the Claude Code binary and drives it from your code. It comes with the built-in tools (read and write files, run commands, search the web), hooks, subagents, MCP connections, permissions, sessions and loading of skills and memory files from .claude/, just as in Claude Code.
When is that worth it? When the agent does not work for you but becomes part of a product or an internal tool: a review agent in your build pipeline, a research agent behind a form, an agent that pre-sorts receipts every night. For languages other than Python and TypeScript, Anthropic names calling Claude Code directly as a subprocess with -p as the alternative.
Two rules from the documentation you should know before you start:
- Sign-in by API key once others use your product. Unless previously approved, Anthropic does not allow third-party developers to offer claude.ai login or subscription rate limits in their products, including agents built on the SDK. So for a product you offer to others, you need an API key. If you only use the SDK yourself while signed in with your Claude plan, Anthropic says it draws on your plan limits. How to create and protect a key is covered in Claude API Key: Setup, Pricing, First Call.
- Your own name. Your product must not be called “Claude Code” or look like it. Allowed are, for example, “Claude Agent” or “Your Name Powered by Claude”.
My team does not run on the SDK but directly in Claude Code on a subscription, because I work with the employees myself. The SDK only becomes interesting when an agent is meant to run without you inside a product.
Claude Managed Agents: Anthropic runs the loop
Managed Agents is the route for anyone who wants an agent but does not want to build an agent loop, a sandbox or tool execution themselves. Anthropic describes it as a pre-built, configurable agent harness on managed infrastructure, meant for long-running and asynchronous work. The agent can read files, run commands, browse the web and run code there. The setup rests on four concepts:
- Agent: model, system prompt, tools, MCP servers and skills.
- Environment: where sessions run, in Anthropic's cloud sandbox or in a sandbox on your own infrastructure.
- Session: a running agent performing a specific task.
- Events: the messages between your application and the agent.
According to the documentation, Managed Agents suits tasks that run for minutes or hours, sessions with a persistent file system and recurring runs on a schedule. Two limitations are stated explicitly: Managed Agents is in beta, and every request needs a dedicated beta header (the SDKs set it automatically). And because sessions store their state on Anthropic's servers, Managed Agents is currently not eligible for Zero Data Retention. You can delete sessions and uploaded files at any time through the API. If you work with customer data, clarify this before the first test; the basics are in Claude Privacy and GDPR: What Applies.
Billing has two parts: tokens at the normal model price and runtime at 0.08 dollars per session-hour, counted only while the session is actually running. Anthropic's own worked example: one hour with Opus 5.5, 50,000 input tokens and 15,000 output tokens, costs 0.58 dollars; with caching for 40,000 of the input tokens, 0.43 dollars. The batch discount does not apply to Managed Agents. What the models cost per token is in Claude Models: Opus, Sonnet, Haiku, Fable.
Messages API: build the loop yourself
The fourth route is the oldest: you send requests to the Messages API, give Claude your own tools and write the loop yourself that executes tool calls and returns the results. The SDK for your language includes a beta “tool runner” that takes over this loop. The upside is full control; the price is building yourself what Claude Code and Managed Agents bring ready-made. I would only recommend it if you know exactly why the ready-made variants fall short.
Claude Duet: the ready-made agent for end users
Duet is the opposite of the developer offerings: its own app in which Claude acts as a proactive agent, keeps an eye on your connected accounts, sends you a brief in the morning and does tasks on its own computer in the cloud. Before sending, paying, signing or deleting, Duet asks. Since October 21, 2026, Duet has been available as a research preview by waitlist, only for Pro and Max subscribers in the US, and it runs on Opus 5.5. What exactly Duet can do and what that means for users outside the US is covered in Claude Duet: What Anthropic's Agent Does.
Skills as the shared building block
One building block runs through almost every route: skills. The Agent SDK loads them from .claude/ just like Claude Code, and Managed Agents includes them in the agent definition. On October 22, 2026, Anthropic handed the open Agent Skills standard to the Agentic AI Foundation under the Linux Foundation; according to Anthropic, more than 70 agent platforms have integrated it, including ChatGPT, Cursor, GitHub Copilot and VS Code. For you that means: a well-written work instruction as a skill is the investment you are most likely to carry from one agent route to the next. How to write skills is covered in my article “Claude Skills: How to Create and Use Them”.
Which route fits you
My recommendation, from simple to elaborate:
- You want to hand off work without programming: Cowork in the Claude chat or a subagent in Claude Code. The starting point is Build an AI Agent With Claude: 3 Ways.
- You work with several roles every day: Claude Code with subagents, like my team. What that looks like is shown in My AI Workforce: Fourteen Employees, One Boss.
- The agent should live in your own software: the Agent SDK if you want to run it yourself; Managed Agents if Anthropic should provide sandbox and runtime.
- You need every step under your own control: the Messages API with your own loop.
- You want a personal agent for your inbox and calendar: Duet, as soon as it is available to you.
For routes 3 and 4: if you pay for Max or Team, according to Anthropic the monthly API credit covers the API, the Agent SDK and Managed Agents, but not interactive Claude Code sessions. How much credit each plan gets is in Claude API Key: Setup, Pricing, First Call. All plan prices are in Claude Pricing and Plans Explained.
Frequently asked questions
What are Claude agents?
An umbrella term for the ways in which Claude plans steps and uses tools on its own: Claude Code with subagents, the Claude Agent SDK, Claude Managed Agents, your own loops on the Messages API and the end-user app Claude Duet. There is no single product called “Claude Agents”.
What is the Claude Agent SDK?
A library for Python and TypeScript that brings Claude Code's agent into your own application: the same tools, the same loop, the same context management. For a product you offer to others, the documentation says to sign in with an API key and pay per token. If you only use it yourself while signed in with your Claude plan, Anthropic says it draws on your plan limits.
How much do Claude Managed Agents cost?
Tokens at the normal model price plus 0.08 dollars per session-hour, counted only while the session is running. Anthropic's example for one hour with Opus 5.5 comes to 0.58 dollars. The batch discount does not apply here.
Can I build a Claude agent without programming?
Yes, through Cowork in the Claude apps or as a subagent in Claude Code, which only needs a text file with instructions. The Agent SDK, Managed Agents and the Messages API, on the other hand, require programming skills.
Can I use my Claude subscription for a product built on the Agent SDK?
Not without prior approval from Anthropic. According to the documentation, third-party developers may not offer claude.ai login or subscription limits in their products; sign-in by API key is the intended route.
What is the difference between a subagent and Managed Agents?
A subagent is a helper inside your Claude Code session, on your computer, billed through your subscription or your key. Managed Agents is a standalone agent that Anthropic runs in the cloud and that your application starts and steers through the API.
Where to go from here
Start with the route that builds the least: a subagent for a task that comes up every week. Only once it is clear that the agent belongs in a product or has to run without you is it worth looking at the SDK and Managed Agents. If you still lack the basics for that, begin with Claude Code: A Guide for Beginners. And if you want to see how others build their agents and AI employees, join my German-speaking community Claude Practitioners.
