Claude Managed Agents: Hosted AI Workers for Small Teams

Most small businesses meet AI through a chat window. You type, it answers, you copy the result somewhere else. A different category of tool has been growing alongside: an “agent” that is given a goal, a set of tools, and permission to work through the steps on its own. Building one used to mean wiring together servers, sandboxes and permission systems. In April 2026 Anthropic put a hosted version of that plumbing into public beta, called Claude Managed Agents. This article explains it without the jargon, and then asks whether a small team needs it at all.

What it is

According to a write-up of the launch and to Anthropic’s own platform release notes, Managed Agents is a set of building blocks for running Claude as a long-lived, task-doing worker on Anthropic’s infrastructure rather than your own. You define an agent with instructions, tools and permissions, then start sessions in which it does the work. Each session runs inside its own isolated container. The agent can use built-in tools for running code, browsing the web, and handling files, and it can talk to outside services through the Model Context Protocol (MCP), a standard way for AI tools to connect to other software.

The point is that someone else handles the unglamorous parts: keeping the sandbox safe, storing credentials, and managing state. One write-up describes a credential vault that holds secrets encrypted and hands them to the agent when needed without exposing them in your code or logs.

What has been added since launch

The release notes read like a list of things a business would eventually ask for. A few examples, with dates from the notes:

Memory for agents entered public beta on April 23.
Webhooks arrived on May 6, so other systems can be told when something happens.
Scheduled deployments appeared on June 9, which lets an agent run on a timetable rather than only on request.
Session budgets on August 7: a session that reaches its spending cap pauses instead of continuing.
Permission policies gained an auto option on September 10, where the server evaluates each tool call and allows, denies or pauses it.
OpenTelemetry export (beta) on October 14, for tracing what a session did.

Notice the theme: budgets, scheduling, logging and approvals. Those are the features that make an autonomous tool tolerable to run in a business.

What it might cost

Pricing details in the launch write-up included model token charges, a fee of $0.08 per hour of active runtime (idle time not billed), and $10 per 1,000 web searches. Prices change, and this figure comes from a third-party article, so check the current numbers on Anthropic’s pricing page before relying on it. The structure is still informative: you pay for thinking, for time the agent is actually working, and for particular tools. A task that runs for a few minutes a day is a very different bill from one that loops all afternoon, which is exactly why the budget cap matters.

Who this is really for

Managed Agents is an API product. It is aimed at developers who want to embed an agent in their own product or internal system. A shop owner will not “turn it on” in a settings page. That means the realistic ways a small business benefits are:

A developer or agency builds a narrow agent for you, such as one that reads incoming supplier emails, checks them against an order sheet, and drafts replies for a person to approve.
You use a product that is itself built on top of it, without ever touching the API.

Good uses, and when to skip it

Where it fits

Work that needs judgement across several steps, such as researching a lead, summarising what you found, and filing a note in your CRM.
Jobs that must run on a schedule and leave a trail, like a weekly review of site errors or content gaps.
Tasks that need a safe place to run code or browse, which you would rather not provide on your own machine.

Where it does not

Fixed, repeatable workflows such as “when a form is submitted, add a row and send a confirmation.” A plain automation tool is cheaper, steadier, and easier to explain.
Anything where a mistake is expensive and irreversible. If you cannot describe the approval step, do not automate the task yet.
Teams with no one to look after it. An agent still needs prompts adjusted, logs checked, and budgets reviewed.

Questions to ask before building one

What exactly is the goal, and how would we know it was done well?
Which tools and accounts does it need, and can each be limited to read-only at first?
Who approves actions that touch customers or money?
What is the spending cap per session and per month?
Where will the log of its actions be kept, and who reads it?

If those answers come easily, a pilot is reasonable. If they do not, a simpler automation will probably serve you better for now. We are happy to help scope this kind of work; see our services for how we handle automation and AI projects for small teams.