Claude Managed Agents: What They Are & When to Use Them

A finance team recently automated its monthly close process after losing its controller. A sales team built an agent that takes a company name and a contact, researches the account, and generates a ready-to-send outreach asset. A marketing team cut the cost of building a client-ready slide deck from $30 to $50 down to under $4, without changing the underlying skill files.

All three were built on Claude Managed Agents, a hosted service that Anthropic says gets teams to production 10x faster. Most agent projects stall after the demo works, when the team has to figure out where it runs, how it recovers when a session drops, and how it handles credentials safely. Managed Agents handles all of that.

What Claude Managed Agents actually is

Most people picture an AI agent as a chatbot with extra steps. Claude Managed Agents is a different thing: a hosted service where Anthropic runs the agent on your behalf, start to finish.

The clearest way to understand it is as one rung on a three-rung ladder, based on how much of the work you build yourself versus how much Anthropic handles for you.

Rung one: the Messages API. You build the entire agent loop yourself. You decide what tools it has, you manage the production infrastructure, and you get tokens in, tokens out. Everything else is on you.

Rung two: the Agent SDK. Anthropic ships you the harness, the same engine behind Claude Code, as software you run yourself. You get file access, session primitives, and observability out of the box, but you still own and operate it on your own infrastructure.

Rung three: Claude Managed Agents. Anthropic hosts and runs the entire runtime. The sandbox, the credential handling, the session recovery, the observability dashboard: all of it lives on Anthropic’s servers, not yours.

Each rung trades a little more control for a lot less infrastructure work. Managed Agents is the option built for teams who want the outcome an agent produces, not a new piece of infrastructure to maintain.

Why this is Anthropic’s recommended approach, not just a new feature

There’s a specific engineering reason Anthropic built this, and it’s worth understanding because it explains why a custom harness is a riskier long-term bet than it looks.

Every custom-built agent harness encodes assumptions about how the underlying model behaves. Those assumptions go stale every time the model improves. Anthropic has seen this happen directly: an earlier version of Claude would wrap up tasks prematurely as its context window filled up, a behavior engineers had to build workarounds for. When a newer model no longer had that behavior, the workaround didn’t just become unnecessary. It became dead weight that actively hurt performance and cost.

A team running its own harness has to catch and fix that kind of thing on every model release. Claude Managed Agents is built so the parts that change constantly, the model and the harness underneath it, stay separate from the parts that need to stay stable, like the record of what the agent has done. That separation is why Anthropic recommends this as the default path for internal business process agents, not just one option among several.

When it makes sense to use a managed agent

Six specific problems tend to show up whenever a team tries to build and run agents on their own. Each one is a reason to consider Managed Agents instead of a custom build.

  1. You don’t want to own hosting and scaling. Managing the containers an agent runs in, and making sure they scale under load, is a real infrastructure job. Most teams don’t want to staff that.
  2. The work needs to survive failures. If something goes wrong mid-session, you want the agent to pick back up where it left off, not restart from scratch and lose the progress it made.
  3. The task needs a consistent workspace. An agent working across multiple runs on the same problem needs a stable place to keep its files and its progress. Building that yourself is harder than it sounds.
  4. You need to contain what happens when something breaks. An agent that executes code or calls tools can occasionally do something you didn’t intend. You want that contained, not free to touch everything your application can touch.
  5. Credentials can’t sit next to agent-generated code. If the environment where an agent writes and runs code also holds your API keys and access tokens, you’ve created a security risk that’s hard to fully close off.
  6. You want visibility without building it yourself. Knowing what an agent did, how long it took, and what it cost shouldn’t require standing up your own monitoring system.

If two or more of these show up in a project, that’s a strong signal a managed agent is the better starting point, not a custom harness.

What it takes to do this well

Building a reliable agent on this platform isn’t just a matter of turning it on. A few practices separate the builds that work consistently from the ones that don’t.

Define the outcome, not just the prompt. The most reliable builds give the agent a clear goal and a rubric for success, then let it iterate against that rubric until the work meets the bar. That’s a different discipline than writing a good prompt. It’s closer to writing a job description.

Let the agent improve on its own. One of the more distinctive capabilities here is a self-review process, sometimes called dreaming, where the agent looks back at its own past sessions, notices patterns, and adjusts its own behavior going forward. That reduces the amount of ongoing tuning a team has to do by hand.

Treat security as structural, not procedural. In a well-built managed agent, the environment where the agent executes code never has direct access to credentials. Access happens through a vault and a proxy instead. That’s a meaningfully stronger guarantee than a policy that says “don’t put secrets in the sandbox,” because it removes the possibility entirely rather than relying on someone following a rule.

Where this leaves a business making the call

The pattern across every internal example here is the same. A team had a business process worth automating, chose not to build the underlying infrastructure themselves, and got a working agent faster because of it.

That’s the actual decision in front of most teams evaluating this: not whether AI agents are worth building, but whether it makes sense to build the plumbing underneath one from scratch. For most internal business processes, the answer is increasingly no.

Real pilots are already moving through approval on this approach, which says something about where the pattern is headed. Teams evaluating whether to build an agent on their own infrastructure or on a managed platform are, in effect, choosing between reinventing infrastructure Anthropic has already built and tested, or building on top of it.

Frequently asked questions

What is the difference between Claude Managed Agents and the Claude Agent SDK?

The Agent SDK gives you the same harness Claude Code uses, but you host and run it on your own infrastructure. Claude Managed Agents goes a step further and runs the entire agent, including the sandbox, credentials, and session state, on Anthropic’s servers.

Do Claude Managed Agents cost more than building a custom agent? Not necessarily. Because you’re only billed for active session time and you’re not staffing infrastructure, session recovery, and credential management yourself, the total cost of ownership is often lower than a self-managed build, even before accounting for engineering time.

Is it possible to reduce the cost of a managed agent after it’s built? Yes. Because the skill files that define an agent’s behavior are portable, teams have swapped the underlying model to reduce cost without rebuilding the agent from scratch.

Key Takeaways

  • You don’t want to own hosting and scaling.
  • The work needs to survive failures.
  • The task needs a consistent workspace.
  • You need to contain what happens when something breaks.
  • Credentials can’t sit next to agent-generated code.
author avatar
Vivian Kim VP of Strategic Alliances & Marketing
Scroll to Top