Prompt: A photorealistic editorial wide shot of a Middle Eastern woman in her 30s with closely cropped dark hair and angular features, walking away from camera through a plant-filled biophilic brutalist interior — raw board-formed concrete walls, polished concrete floor, large monstera and trailing pothos filling one side of the frame, dark oak shelving built into the concrete. In a dynamic stride, caught mid-step, body turned slightly away, one hand trailing along a floating dark oak shelf. She wears an oversized cream linen Yohji Yamamoto coat with draped sleeves, wide charcoal trousers, flat leather shoes. Interesting perspective from a low angle behind her. In the middle distance, built into the concrete wall, a single workstation: a thick reclaimed oak slab on matte black steel legs, an open matte black laptop with the screen softly glowing warm from recent activity, a small brass desk lamp left on, a half-drunk glass of water. Nobody is at the workstation — the work is continuing in her absence. Elevated chiaroscuro lighting with lifted blacks — the darkest tones are matte dark slate, never true black. Late afternoon directional light rakes in from a large Crittall-style steel-framed window out of frame to the left, catching the concrete texture, the monstera leaves, and the wool of her coat. Natural skin texture, realistic pores, a small mole on her jawline, no airbrushing. Quiet, contemplative mood. Captured on a Fujifilm GFX 100S medium format with a 24mm lens at f/2.8, Kodak Portra 400 film with lifted shadows and matte finish. Natural film grain. Absolutely no text, no labels, no logos, no readable words anywhere in the image.

Two days ago, something quietly shifted.

You have been working with Claude Code for weeks or months. It runs on your laptop. You open the desktop app, point it at a folder, describe what you want, and Claude builds. If you close the lid, Claude stops. If your battery dies, the session ends. If you go to sleep, so does the work.

On April 8, 2026, Anthropic gave Claude somewhere else to live.

It is called Claude Managed Agents, and it is the biggest structural change to Claude Code since launch. Agents you set up now run on Anthropic’s infrastructure — their servers, their containers, their uptime. Your laptop is optional. You can describe a task, walk away, and come back to finished work.

This guide is about what that actually means for someone who builds creative things with Claude, and what it unlocks that was not possible before.

What a managed agent actually is

Think of Claude Code, up until now, as a brilliant assistant who lives on your computer. It reads your files, runs your browser, watches your terminal. Everything happens on your machine. If you are away from your machine, the assistant is away from work.

A managed agent is that same assistant, but given a workspace of its own — a quiet, clean, persistent room on Anthropic’s servers. You hand it a brief. It goes into its room and works. It has its own computer in there (a fresh, isolated one), its own browser, its own memory of what you have said. You do not need to be watching.

When the agent is done, the session still exists. You can open it the next morning, read what happened, continue the conversation, or tell it to do the next thing. The work waits for you, not the other way round.

That is the shift. Claude is no longer bound to your device.

What this unlocks for creatives

Most articles about this launch are framed for software developers — CI pipelines, production deployments, API integrations. That is real, but it is not the most interesting thing.

Here is what genuinely changes when Claude has its own place to work:

Overnight jobs. You finish a shoot at midnight. Instead of sitting at your desk organising 2,000 photos, you describe the task to a managed agent — sort by date, rename by subject, build a contact sheet, flag the ten best based on sharpness and composition — and go to bed. You wake up and the work is done.

Parallel work while you are on your actual work. You spend Tuesday morning painting. A managed agent spends Tuesday morning building a new page on your portfolio site. Neither of you blocks the other. You are not babysitting a progress bar.

Agents that just keep running. Want an agent that checks your site every day at 9 a.m. and reports any broken links? Or one that watches a folder for new images and generates captions for each? These existed in clunky forms before — now they live in a real, persistent, hosted environment, running whether your laptop is open or not.

Starting a task from anywhere. From your phone. From a cafe laptop. From any browser. You do not need the desktop app. You do not need your files. The agent has its own files now.

This is the rhythm Claude Code was always pointing toward, but could never fully reach. It has reached it now.

The four pieces

Managed agents are made of four simple concepts. Knowing the names makes everything else easier to follow.

Agent. The recipe. Its system prompt, the tools it can use, any skills it knows, and what it is for. Think of this as the personality and job description. You create it once.

Environment. The workspace — an isolated container with a fresh computer inside. Agents run in environments. You can configure whether the environment has internet access, what files it starts with, and how it is locked down.

Session. A running instance of an agent inside an environment. This is where the work actually happens. Sessions are persistent — you can close them, come back hours later, and pick up where you left off. The conversation history, the files, the state of the machine — all still there.

Events. The stream of messages that come out of a session as it runs. Every time the agent thinks, uses a tool, or speaks, it emits an event. You can watch this stream live, or check in later and read what happened.

That is the full vocabulary. Agent → Environment → Session → Events. Everything else is detail.

Three ways to set one up

You do not need to learn a new tool to use managed agents. They are available from three places, and you can pick whichever feels most natural.

1. The Claude console

The point-and-click option. Go to the Claude platform in your browser, open the console, and create an agent through a visual form. You give it a name, a system prompt, pick its tools, and launch. This is the easiest way to try it the first time.

2. Claude Code itself

The most seamless option for anyone already using the desktop app. Claude Code can create managed agents directly — you describe what you want the agent to do without leaving your current session, and your local Claude Code session and the managed agent show up in the same place. For someone already comfortable with Claude Code, this is where most of this will happen day to day.

3. The ant CLI

A new tool that shipped the same day. ant is Anthropic’s first official command-line client for the Claude API — you install it once (brew install anthropics/tap/ant on a Mac) and it gives you a clean way to create agents, sessions, and environments from the terminal, using simple YAML files instead of code.

This is the option to reach for when you want to version-control your agents the same way you version-control your website — write the agent config in a YAML file, keep it in git, deploy it with one command. If you do not work with the terminal, ignore this one for now. The first two options are enough.

What the agent can do inside its room

Every managed agent comes with the same basic toolset out of the box:

  • Bash. It has a terminal. It can run any command that a normal computer can.
  • File operations. It can create, read, edit, move, and search files inside its environment.
  • Web. It can browse the internet, search the web, and fetch pages.
  • MCP servers. The same external tools you already connect to Claude Code — Figma, Notion, Stripe, your own databases — can be connected to a managed agent. Your stack travels with it.
  • Skills. The skills you have built for Claude Code can be loaded into a managed agent. A /design-check skill on your laptop works the same way when it is running on Anthropic’s servers.

The environment is sandboxed. It cannot touch your actual computer. It cannot read files outside its own workspace. You decide how locked-down the network is — fully open, or restricted to specific domains only.

What it costs (and the fine print)

Managed agents use standard Claude API rates for the tokens the agent consumes — nothing new there. On top of that there is a charge for the time the agent’s environment is actively running. The exact figure is not on the official overview page at the time of writing, so rather than guess, check the Anthropic pricing page before you spin up something long-running.

A few practical notes:

  • Managed agents are open to all Claude API accounts by default. No waitlist for the core product.
  • Three advanced features — multi-agent orchestration, custom outcomes, and long-term memory — are in gated research preview. You can apply for access if you need them.
  • The whole thing is in public beta. The product will change. Features will land. Things you read today may be different in a month. Check the official docs before building anything important around it.

Should you use this yet?

Honest answer: not for everything.

If your day is mostly describing a change and watching Claude make it on your laptop, the desktop app is still the right tool. It is faster, simpler, and free of the mental overhead of thinking about remote environments.

Managed agents are the right tool when:

  • You want work to happen while you are not at your computer.
  • You have a task that takes long enough to be annoying to watch — a big photo sort, a site-wide audit, a generative experiment that runs for an hour.
  • You want an agent that lives somewhere permanent and can be triggered on a schedule.
  • You want to hand a task to Claude from your phone, a friend’s laptop, or anywhere that is not your main setup.

For most creatives, the answer will be: interesting, filing it away, will reach for it when the right task arrives. That is the right answer. The feature is new. The patterns for using it well are still being written — by everyone, including Anthropic.

A small experiment to try this week

If you want to feel the difference without committing to anything, try this:

  1. Open Claude Code on your desktop.
  2. Type /agents and look at the new tab for managed agents.
  3. Create a simple one — give it the system prompt “You are a careful proofreader. Read any document you are given and return a list of typos, awkward phrasing, and unclear sentences.”
  4. Launch a session. Paste in a few paragraphs of your own writing.
  5. Close your laptop. Walk away. Come back later. The session is still there. The work is done.

That small loop — set, close, return — is the feel of the whole thing. Once you have done it once, you will know in which corners of your week it belongs.

The bigger picture

Everything about Claude Code so far has been about making a single person, on a single machine, more capable. Managed agents start the next chapter: making work itself more capable, regardless of where anyone happens to be standing.

For someone who paints in the morning, shoots in the afternoon, and only touches the laptop for an hour after dinner, this is not an abstraction. It is permission to stop rearranging your day around your software.

Claude learned to work without you. That is a good thing for everyone it works for.

Art & Algorithms publishes guides, tutorials, and prompt packs at the intersection of art and code. Subscribe for the full archive.