Anthropic · April 2026 · Feature Briefing

Claude Managed Agents

A hosted infrastructure for running Claude as an autonomous, long-horizon agent — without building your own scaffolding.

Beta Long-running tasks Multi-tool Stateful sessions MCP support

What is Claude Managed Agents?

You've been using Claude Code to run tasks — writing files, executing bash commands, searching the web. That's already powerful. But imagine delegating an entire multi-hour project to Claude, walking away, and coming back to results — without building any of the scaffolding yourself.

Managed Agents is Anthropic's new hosted service that runs Claude as a fully autonomous agent inside a secure cloud environment. You describe the task, Claude takes over — running commands, reading files, browsing the web, writing code — and reports back when done.

The One-Liner

It's a cloud platform where you create your own Claude version, with the infrastructure handled — and you call it through an API.

Analogy

Think of it as the difference between giving someone a one-off instruction vs. hiring a contractor who has their own office, tools, and can work overnight on your project. You just call them when you need something done.

Do you need an IDE?

Yes — to call the API you need to write code somewhere. VS Code or any editor works. The agent runs in Anthropic's cloud; your script is just the trigger that lives on your machine.

The core idea: brain vs. hands

The key innovation is a separation engineers call decoupling the brain from the hands. Before this, Claude's thinking and actions lived in the same box. If it crashed, everything was lost. Now they're separate — which changes what's possible.

🧠
Agent = The Worker

The model, its instructions, and which tools it can use. Define this once and reuse it across any number of tasks.

🏭
Environment = The Workshop

A cloud container with pre-installed software — Python, Node.js, Go, whatever the task needs. Like giving the worker a fully equipped factory before they arrive.

📋
Session = The Project File

A running instance of the agent working on a specific task. Everything that happens gets logged here — even if Claude crashes and restarts, the log survives.

💬
Events = Notes on the Desk

The back-and-forth between you and Claude mid-task. Send a message to steer, pause, or redirect — like leaving a note for a contractor while they work.

Why does this matter?

The session log lives outside the agent. If Claude's process crashes, a new one picks up from the last event — like a GPS that recalculates when you take a wrong turn. Nothing is lost.

The Operating System Analogy

Anthropic compares this to how OS abstractions work. A read() command works whether it's reading from a 1970s disk or a modern SSD. The interface outlasts the hardware. Managed Agents is built the same way: stable interfaces on top of an evolving implementation.

What Claude gets out of the box

When you spin up a session, Claude has a full toolkit available — no manual wiring needed.

bash Run shell commands

Execute scripts, install packages, run test suites — anything you'd do in a terminal.

file ops Read, write, edit files

Full filesystem access inside the container. Claude can create, modify, and search files autonomously.

web search Search the web

Look up documentation, find libraries, research errors — in real time during the task.

web fetch Retrieve content from URLs

Pull full web pages, JSON APIs, changelogs — anything accessible via HTTP.

MCP Connect external tools

GitHub, Slack, databases — anything that speaks MCP. Credentials are stored in a separate vault; Claude can use them but can't access them directly.

Security Note

Your credentials never live inside the sandbox. They live in a separate vault. Claude can use them, but can't read them — like a hotel safe a valet can open for you but can't pocket.

How to use it

Since you already use Claude Code, think of Managed Agents as the API-level version — define the worker, give them a workspace, start a task, and let it run.

1
Create an Agent

Define which Claude model to use, write a system prompt, and specify which tools it can access. Do this once — reuse the agent ID across many sessions.

2
Create an Environment

Configure the cloud container: packages pre-installed, network access rules, files to mount. Like setting up a workspace before the worker arrives.

3
Start a Session

Launch a session linking your agent to your environment. Each session has its own persisted event log.

4
Send a Task and Stream Results

Send your task as a user message. Claude starts working autonomously and responses stream back in real time via server-sent events.

5
Steer or Interrupt

Mid-task, send additional messages to redirect Claude, ask for a status, or stop it. The full event history is always retrievable.

vs. Claude Code

Claude Code is an excellent harness that Anthropic themselves use. Managed Agents is the infrastructure layer that can run Claude Code, a custom harness, or a task-specific agent — all through the same stable interfaces. It's not a replacement; it's the platform underneath.

End-to-end: automated dependency audit

You maintain a Node.js service. Every month you want someone to check for outdated dependencies, run the test suite, and open a GitHub PR — while you sleep.

🏗️ Step 1 — Define the Agent
system: 'Check outdated npm packages, run tests, open a PR if tests pass.'

Write this once. The agent is reusable across every future monthly run.

📦 Step 2 — Configure the Environment
packages: [node@20, npm, git] network: allow github.com, registry.npmjs.org mount: /repo -> your GitHub repository clone

The container comes pre-loaded with Node and Git. Your repo is mounted and ready.

▶️ Step 3 — Start the Session

One API call starts the session. The agent wakes up inside the container and begins working immediately.

⚙️ Step 4 — Claude Works Autonomously
-> bash: npm outdated -> bash: npm update -> bash: npm test -> web_fetch: github.com/api (open PR) -> emit: PR #47 opened — bump lodash 4.17.19 → 4.17.21

Claude runs commands, reads the output, handles failures, and opens the PR — without you touching anything.

📬 Step 5 — You Get a Report

The session streams back a summary. Query the full event log anytime to see exactly what Claude did, step by step.

The Payoff

The same agent definition handles this task every month. If Claude's process crashes mid-run, a new one picks up from the last session event — nothing is lost. As Claude models improve, the agent gets smarter without any changes on your end.

Code Examples

Same task, three languages. The pattern is always the same: create agent → create environment → start session → send task → stream results.

Setup — run in your terminal first
What Happens When You Run This
1.

Anthropic spins up a cloud container with npm + git pre-installed

2.

Claude reads your system prompt and picks up the task

3.

It runs bash commands, checks results, handles failures autonomously

4.

Your terminal streams back each step in real time

5.

When done, the session log is saved — query it anytime