Claude Managed Agents is Anthropic's service for running an artificial intelligence agent on the vendor's own infrastructure. You describe the agent, its model, its tools and its instructions. The platform supplies the rest: the execution loop, an isolated sandbox, session persistence and readable traces. The service has been in public beta since April 2026.
The promise is clear for a technical team in a small or mid-sized company. It removes several weeks of plumbing between an agent that works on a laptop and an agent that runs in production. The question is what that plumbing really cost, and what the platform still does not cover.
What is Claude Managed Agents, exactly?
Claude Managed Agents is a pre-built agent harness that runs in infrastructure managed by Anthropic. The official documentation presents it as the second way to build with Claude, alongside the Messages API. The Messages API gives direct model access and leaves you to write your own loop. Managed Agents supplies the loop, the tool execution and the runtime.
Inside that harness, Claude reads files, runs shell commands, browses the web and executes code in an isolated container. Prompt caching and context compaction are built in, two optimisations a team usually spends weeks getting right.
Four objects describe the whole model
The vocabulary fits into four notions, which is good news for a team new to the subject. The agent carries the model, the instructions, the tools and the MCP servers. The environment says where the code runs. The session is one running execution on a given task. Events are the messages exchanged between your application and the agent.
What you no longer have to build
Tandem has shipped agents to production across more than 50 companies, and the same finding comes back on every engagement. The agent code takes a fraction of the time. The rest goes into the sandbox, recovery after incidents, state storage, secret rotation and observability. Managed Agents moves that load to the vendor.
Scheduled deployments and credential vaults
Two pieces shipped in June 2026 deserve a director's attention more than a developer's. Scheduled deployments start an agent on a cron-style recurrence, each run opening a fresh session. Vaults store third-party credentials. Anthropic writes that never sees your key, the sandbox holding only a placeholder, with the real value attached at the network boundary for approved domains only.
That second piece answers an objection IT departments raise in every workshop. An autonomous agent holding the CRM and mailbox credentials in clear text is a risk nobody wants to carry. Separating the agent from its secrets changes the conversation.
What does a hosted agent cost at Anthropic?
The bill has two lines, and Anthropic's pricing page states them without ambiguity. Tokens are billed at the usual model rate. Session runtime is added, at around 7 cents per hour, measured to the millisecond and counted only while the session is running. A session waiting for an answer or an approval costs nothing.
The order of magnitude is worth saying out loud. An agent running without interruption for a whole month would cost around 50 euros of session runtime. A dedicated server, its monitoring and the engineering time around it go well beyond that. The line that decides the bill remains token consumption, which a poorly scoped agent drives up fast.
Does it replace n8n, Make or Dust?
No, and the confusion deserves a direct answer because it has come up in almost every executive committee since the spring. An agent harness and an automation platform solve different problems. The first runs autonomous reasoning. The second runs a business process that non-technical teams must be able to read, restart and correct.
I made the argument in February 2026 in a post about infrastructure: when an agent breaks at three in the morning, you do not want a terminal, you want readable logs, automatic retries and alerts. That is exactly the gap Managed Agents addresses, and the platform fills a real part of it with its session console and its webhooks.
Only a part of it. Anthropic's console is restricted to developer and admin roles on the account. An order management lead or a financial controller will not go in, and will not restart a failed run. What those teams need is a business interface, and it still has to be built, either by you or on a platform that already provides one.
| Need | The right tool | Why |
|---|---|---|
| Long autonomous task, launched from code | Claude Managed Agents | Sandbox, persisted state and recovery provided |
| Business process owned by non-technical teams | n8n or an equivalent platform | Readable canvas, visual retries and alerts |
| Internal assistants plugged into company data | Dust or equivalent | Access and workspace management per team |
| Full control over execution and context | Self-hosted Agent SDK | No black box, but all the infrastructure to carry |
The reasoning behind that choice matters more than the choice itself. I covered it in a video on the layers of automation, where I argue that the agent is rarely the first step. In many cases a simple automation or a well-configured assistant is enough, and costs far less to maintain. My review of n8n and the one on the Dust platform complete the picture.
Where does your data live with a hosted agent?
This is the question that decides deployment in a European company, and Anthropic's answer has two levels. The first is the self-hosted sandbox. The second is a limit worth knowing before committing to a project.
With a self-hosted sandbox, tool execution leaves Anthropic's infrastructure. A program you run on your own machines claims the work, runs the tool calls locally and posts the results back. Claude keeps reasoning on the vendor side and decides which tool comes next. Tool inputs and outputs therefore still flow through Anthropic's control plane, and the documentation states it plainly.
Beyond that, geographic pinning of inference is set at the agent level, with a per-session override. The security rules that apply to any autonomous agent do not change, and I gathered them in our guide on agent security.
Multi-agent and research preview: where is the beta?
Multiagent orchestration has left research preview. It is documented as beta, under the same header as the rest of the service. An agent can delegate to subagents whose reports it reads, or write a dynamic workflow, a program that launches many agents in the background and combines their results.
Two pieces remain in a narrower research preview and require an explicit access request: MCP tunnels and the dreaming mechanism. Any write-up presenting the whole service as unreservedly available is wrong on those two points.
Should you run your AI agents at Anthropic?
Yes, if your team already writes code and the agent works long stretches on technical tasks. The gain is immediate, session runtime amounts to a few tens of euros a month, and you recover weeks of engineering. The self-hosted sandbox lifts the data residency objection in most configurations.
No, if the goal is for a business team to own a process. In that case the harness solves the cheapest problem and leaves the most expensive one untouched, the interface and the supervision. Across around 100 engagements, the gap between an agent that works in a demo and an agent teams use daily has always come down to that.
The right sequence has not changed with this release. You frame the process, you measure what it costs today, then you pick the lowest layer of automation that solves the problem. Managed Agents is an excellent tool for the top layer. It does not exempt you from checking that you need it.
