Claude Managed Agents is Anthropic's hosted service for running Claude as an agent: you define the job, the tools and what the agent can reach, and Anthropic runs the agent loop, the sandbox, memory and the schedule. You need it for long-running work that belongs to a system rather than one person, and for one person's recurring job a Cowork scheduled task already does it. It isn't eligible for Zero Data Retention or a HIPAA BAA, and running the sandbox on your own infrastructure doesn't change that.

Clients have started asking me about it. The question usually comes from someone who read the launch post and wants to know whether it's the thing to buy. Most of what's been written is for developers, and this is the version for the person who signs off.

Anthropic launched it on April 8, 2026 as "a suite of composable APIs for building and deploying cloud-hosted agents at scale." It's in beta, and it's enabled by default for every Claude API account. If you're evaluating agents, it deserves a serious look: it takes away most of the infrastructure that made agents slow to build.

What Anthropic runs for you

An agent needs a loop that decides what to do next, a safe place to run commands and work on files, somewhere to keep its history, and a way to hold credentials without exposing them. Teams used to build all of that themselves. Anthropic's docs put it plainly: "Instead of building your own agent loop, tool execution, and runtime, you get a fully managed environment where Claude can read files, run commands, browse the web, and run code securely."

Your side is the definition of the job: the model, the instructions, the tools the agent may use, and the systems it can reach through MCP servers, the open standard Claude's connectors use. Anthropic's product page lists the rest: memory that carries from one session to the next, a lead agent that hands work to specialists, a grader that sends work back until it meets your bar, vaults that keep credentials out of the agent's workspace, and session tracing. Scheduled runs arrived on June 9.

The work runs in a sandbox. By default that's a fresh container Anthropic hosts for each session. If the agent's files have to stay on your own infrastructure, you can run the sandbox yourself, and Claude still runs on Anthropic's side.

When you need it, and when a scheduled task already does the job

Anthropic says the product is best for "Long-running tasks and asynchronous work", meaning work nobody sits and watches. Its own examples include agents that plug into Slack and Teams and return a spreadsheet or a deck, and a weekly compliance scan that runs on a schedule.

Before you build one, check whether you need one. Cowork scheduled tasks come with every paid Claude plan, and Anthropic's help page (read November 2) says they "run remotely, so they run on their cadence even when your computer is asleep or the Claude Desktop app is closed." A morning inbox pass, a weekly report from Drive, a daily Slack summary: that's what scheduled tasks are for, and it's what Lone Star's legal assistant runs every weekday morning. No API account, no code.

If one person needs the same job done every Monday, that's a scheduled task. If the job belongs to a system, that's where an agent earns its place.

Managed Agents is the step up when the agent belongs to a system rather than a person. It serves a whole team or your customers, your own software triggers it, it needs its own credentials and permissions, or it runs long enough that you want a record of every step and a spending cap on each session. It's configured through the Claude Console and the API, not the Claude app, so it's a build, not a setting.

There's a third option. Anthropic's Agent SDK gives developers the same kind of agent loop as a library inside an application you operate. Anthropic's own comparison puts the difference in who runs the agent: with the SDK you do, and with Managed Agents Anthropic hosts it.

The data question to settle first

This is the line most write-ups leave out. Anthropic's documentation says Managed Agents stores conversation history, sandbox state and outputs on its side, and "Because of this, Managed Agents is not currently eligible for Zero Data Retention or HIPAA Business Associate Agreement (BAA) coverage."

Running the sandbox on your own infrastructure doesn't change that. The same page's data-handling table marks both "Not eligible" and says it "Applies to all Managed Agents sub-features, including Self-hosted sandboxes." The environment guide gives the reason: "the conversation history and tool results go to the Managed Agents service, which Anthropic operates."

The files stay on your machines. What the agent reads from them and sends back does not.

You can delete sessions and uploaded files through the API at any time. Whether that's enough for your data is a question for your own counsel or compliance lead, and it's better answered before a build than after one. If your Claude contract depends on zero data retention, or the agent would touch protected health information, ask that question first. Where your files go in Claude covers the rest of the product.

Before you start

It's billed on consumption, tokens plus session runtime, against a Claude API account that's separate from your team's seats. Anthropic's pricing page has the rates. Decide whose account the agent runs under before anyone builds it, because that account gets the bill and holds the history.

It's in beta, and Anthropic ships new features almost every day, so quote the current docs rather than the launch post. Two features the April post put behind an access request are now listed as built in. If customers will see the agent, Anthropic's branding guidelines let you call it "Claude Agent" or "Powered by Claude," never "Claude Code" or "Claude Cowork."

The quickest way to understand it is to try it. The Claude Console starts you from a template, such as a research agent, a support agent or a contract tracker, with the tools and permissions already set.

Managed Agents is the right home for an agent that belongs to a system, and Anthropic has done the hard infrastructure work for you. For one person's Monday report, a scheduled task already does the job. Settle the data question first, and what's left is a build.

Questions

What is Claude Managed Agents?

Anthropic's hosted service for running Claude as an agent that works through a multi-step task on its own, with tools, for minutes or hours. You define the job, the tools and what the agent can reach, and Anthropic runs the agent loop, the sandbox, memory and scheduled runs. It's in beta and enabled by default on every Claude API account.

How is Managed Agents different from the Claude Agent SDK?

Anthropic's own comparison puts the difference in who runs the agent. The Agent SDK gives developers the agent loop as a library inside an application you operate, so you run it. With Managed Agents, Anthropic hosts it.

Do I need Managed Agents for a recurring task?

Often not. Cowork scheduled tasks come with every paid Claude plan, and Anthropic's help page (read November 2, 2026) says they run remotely, even when your computer is asleep or the desktop app is closed. Managed Agents is the step up when the agent belongs to a system: it serves a team or your customers, your own software triggers it, or it needs its own credentials.

Is Claude Managed Agents eligible for zero data retention or HIPAA?

No. Anthropic's documentation marks Managed Agents not eligible for Zero Data Retention or HIPAA Business Associate Agreement coverage, including with a self-hosted sandbox, because conversation history and tool results always go to Anthropic's service. You can delete sessions and uploaded files through the API. Whether that's enough for your data is a question for your own counsel or compliance lead.

How is Claude Managed Agents billed?

On consumption: tokens plus session runtime, against a Claude API account that's separate from your team's seats. Anthropic's pricing page lists the rates.