← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Claude Managed
Agents

You tell Claude what you want an AI agent to do, and Anthropic runs it, hosts it and keeps it going on its own servers. Here’s what that means and how to use it.

On April 8, 2026, Anthropic launched Claude Managed Agents in public beta on the Claude Platform, its developer platform (announcement). It’s still in beta today, and it’s on by default for every API account.

Before this, building an agent that works on its own meant building everything around the model yourself. The loop that keeps it working, running its tools, handling errors, the servers. Anthropic says that used to take months.

Now Anthropic handles that part. You define the agent and give it a task. It runs in a secure cloud workspace, and if a task takes hours, it keeps going. Notion, Asana, Rakuten and Sentry were among the first companies building on it.

Think of it this way

Before, you had to build the car, the road and the gas station. Now you just say where you want to go.

The 4 Building Blocks

Once these click, everything else does too.

01

Agent

The blueprint. The model, its instructions, its tools, any MCP servers and skills. Create it once and reuse it.

02

Environment

The workspace. A secure cloud sandbox where it runs code, reads files and reaches the internet, with you controlling which sites it can reach. Companies can also run it on their own servers.

03

Session

One running job. Give the agent a task and a session starts. It keeps going until the work is done, even if that takes hours.

04

Events

The back-and-forth. You send messages, it streams back results and status updates. You can steer it or interrupt it mid-task.

Why the design matters

The session is a saved log of everything that happened, kept separate from the agent and its workspace. If the agent crashes, a new one reads the log and picks up where it left off. Anthropic calls this decoupling the brain from the hands. (Anthropic Engineering)

What It Can Do Out Of The Box

Every agent gets 8 built-in tools, all on by default. You can switch any of them off.

  1. Bash runs commands, like installing a package or running a script.
  2. Read, Write and Edit work with files in its workspace.
  3. Glob and Grep find files and search inside them.
  4. Web search and web fetch pull in live information, and you can limit them to certain sites.

On top of that you can add custom tools (the agent asks your code to do something specific) and MCP servers (connections to outside apps, databases and internal tools).

In practice

One agent can research a topic online, run a Python script on a spreadsheet, pull data from your company’s API and write the report, all in one session, on its own. It does the work instead of answering questions about it.

Set Up Your First Agent

You need a Claude API account. Go to the Claude Console, sign up, add a payment method and create an API key. Then pick a path.

01

Let Claude Code do it

The easiest path. In Claude Code, run the command below and it walks you through setup with Anthropic’s built-in Claude API skill.

/claude-api managed-agents-onboard

02

Use the ant CLI

On a Mac, install it with brew install anthropics/tap/ant. You describe the agent in a short markdown file and the workspace in a YAML file, then run ant apply on both.

03

Call the API

For full control, use /v1/agents, /v1/environments and /v1/sessions, then send events and stream the results back. SDKs exist for Python, TypeScript, Java, Go, C#, Ruby and PHP, and they add the required beta header (managed-agents-2026-04-01) for you. (Quickstart)

Not sure what your agent should look like? This prompt designs it with you and writes the files.

Design My First Managed Agent
You're a senior engineer who builds agents on Claude Managed Agents. Help me design my first one and hand me files I can deploy. The job I want the agent to do: [Describe the task in plain words, e.g. "Every Monday, research my 5 competitors' pricing pages and write a one-page summary of what changed."] What it needs to read or reach: [Websites, files I'll upload, APIs, MCP servers, or "just the web"] What a great result looks like: [The output format, length, where it should land] Steps: 1. Ask me up to 3 questions if anything above is unclear. Then stop and wait. 2. Recommend a model (cheapest one that will do the job well) and explain why in one line. 3. Write the agent file for `ant apply`: a markdown file with YAML frontmatter (name, model, tools) and the system prompt as the body. Use the `agent_toolset_20260401` toolset and turn off any tools the job doesn't need. 4. Write the environment.yaml with networking set to `limited` and only the hosts the agent actually needs in `allowed_hosts`. 5. Write the first user message I should send to start a session. 6. Estimate the cost of one run (tokens plus $0.08 per running session-hour) and say what would make it more expensive. Guardrails: - Least access. No tool, host or permission the job doesn't need. - Anything that sends, deletes, buys or publishes needs my confirmation in the permission policy. - Don't invent API fields. If you're not sure a field exists, say so and point me to the Managed Agents docs.

Models

It works with every current Claude model, including Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1. Start with the cheapest one that does the job. (Docs)

What People Build
  1. Data processing. Upload a CSV and it cleans it, finds trends, builds charts and writes the summary.
  2. Research. Give it a topic and it searches, reads, cross-checks and writes a brief with sources.
  3. Code. Describe a tool and it writes it, tests it and fixes it.
  4. Content. Feed it raw notes and it turns them into posts, captions and newsletters.
  5. Agent teams. A lead agent hands research, writing and review to specialist agents.
  6. Scheduled jobs. Run an agent on a repeating schedule with scheduled deployments.

Some pieces, like agent teams and self-grading, started out as research previews. Here’s what Dreaming, Multiagent and Outcomes do.

What It Costs

You pay for two things: tokens at normal API rates, and $0.08 per session-hour while the agent is actively running. Time it spends idle, waiting for you, doesn’t count. Web searches inside a session cost $10 per 1,000. (Pricing)

Sonnet 5.5 costs $2 per million input tokens and $10 per million output. Opus 5.5 is $4 in and $20 out.

Real example

Anthropic’s own example: a 1-hour Opus 5.5 session with 50,000 input tokens and 15,000 output tokens costs about $0.58. That’s an hour of the agent working on its own for under a dollar.

Which Option Fits

01

Managed Agents

Anthropic runs the loop and the workspace. Best for long-running jobs and when you don’t want to manage servers. Sessions save their own history.

02

The regular Claude API

You build the loop yourself. Best when you want control over every single step.

03

The Claude Agent SDK

A ready-made agent loop you run on your own machine or servers. Great for local automation and CI/CD.

Why This Matters

At first only engineers could build agents. Then technical founders. Now Anthropic handles the hardest parts, and the bar keeps dropping every month.

You don’t need to understand containers to get ahead here. You need to know what’s possible and how to think about automation. That skill transfers to every tool that comes next.

The big picture

The gap between "person who uses AI" and "person who has AI working for them" is closing fast. Managed Agents closed it a little more.