Here’s a question nobody asks until it’s 2 a.m. and the agent stopped working: who’s watching the machine?
When I set up my AI chief of staff, I chose to run it on a dedicated Mac mini. That’s a good setup for me. It also means the machine, its updates, its power and its access are mine to manage.
Not every team wants that job. That’s where Claude Managed Agents comes in.
What Claude Managed Agents is
As of October 2026, Claude Managed Agents is an Anthropic-hosted agent harness, and it’s in beta. A harness is everything around the model that makes it an agent: the loop, the tools, the environment it works in. With Managed Agents, Anthropic runs that loop for you.
You still design the agent. Anthropic hosts it. If you want the side-by-side with every other way to build on Claude, start at the Claude agent overview.
How Managed Agents works, in four moves
1. Define the agent
You describe the agent: which Claude model it uses, its system prompt (the standing instructions it always reads), the tools it can use, the MCP servers it connects to, and the skills it can load. MCP is the open standard for plugging agents into tools and data; I cover it on what MCP is. Skills are folders of instructions loaded only when a job needs them; see Claude skills.
2. Choose the environment
The environment is where the agent does its work. You pick Anthropic’s cloud sandbox or a self-hosted environment. A sandbox is a walled-off workspace, so the agent’s activity stays contained.
3. Run sessions
Each run of the agent is a session. Sessions are stateful, meaning the agent keeps track of the work within that session, and they’re stored server-side. You can delete them.
4. Schedule and coordinate
You can schedule recurring runs, the morning report or the weekly cleanup, and coordinate multiple agents working together. That second part matters once you move from one agent to a team, the pattern I lay out on multi-agent systems.
For the exact setup, the Managed Agents overview in Anthropic’s docs is the source of truth. Because it’s in beta, expect details to change.
Managed Agents vs running your own machine
| Question | Your own machine (my setup) | Claude Managed Agents |
|---|---|---|
| Who runs the agent loop? | You, through Claude Code | Anthropic |
| Where does the work happen? | A computer you own and maintain | Anthropic’s cloud sandbox or self-hosted |
| Scheduling | Headless runs or Claude Code routines | Scheduled recurring runs |
| Several agents | Subagents and your own coordination | Built-in multi-agent coordination |
| What you maintain | Hardware, updates, access, uptime | The agent definition, data rules and review |
| Maturity | Established tool | Beta as of October 2026 |
When Managed Agents beats running your own machine
It fits when:
- Nobody on your team wants to own a computer that has to stay on and stay secure.
- You want agents running in a contained sandbox instead of on a machine with other things on it.
- You’re coordinating several agents and want that handled by the platform.
- You have a developer or technical lead to define the agent and wire the tools.
Your own machine may fit better when:
- You want the agent working inside local files and apps on that machine.
- You’re comfortable maintaining hardware and prefer that control.
- You’d rather wait for a product to leave beta before you put core operations on it.
Data and retention: decide before the first session
Sessions are stored server-side until you delete them. That’s useful for continuity, and it’s also data you’re responsible for. Before you run anything:
- Write down what data the agent may see, and what it may never see.
- Keep client and sensitive data out unless your settings and AI policy allow it.
- Decide how long sessions should live, and make deleting finished ones a routine.
- Give each agent least-privilege access: read-only first, write access one tool at a time.
- Name one human owner who reads the logs and approves anything irreversible.
Hosted or not, outreach rules follow the agent. Under the FCC’s February 2024 decision, an AI voice is an artificial voice for TCPA purposes, so calls need proper consent; texts need consent and A2P 10DLC registration; Do Not Call applies to both. Disclose the AI and get counsel’s review before customers hear from it.
What to hand your technical lead
Bring a one-page brief: the job and its definition of done, the tools and access levels, the schedule, the guardrails and what needs approval, and the data rules including session deletion. That’s the same discipline as any agent build; how to build an AI agent walks the full order.
Your next step
Decide who owns the machine question in your business. If the honest answer is “nobody,” Managed Agents belongs on your shortlist; have your technical lead read the docs and test one low-risk job. If you want hosted or custom agents built for your company, my AI company ClientPro does that work. If you want an AI chief of staff installed with you, see how Chief works.
Questions people ask
What are Claude Managed Agents?
As of October 2026, Claude Managed Agents is an Anthropic-hosted agent harness in beta. You define the agent — model, system prompt, tools, MCP servers and skills — pick an environment, and run sessions. It can schedule recurring runs and coordinate multiple agents.
Where do Claude Managed Agents run?
You choose an environment: Anthropic’s cloud sandbox or a self-hosted environment. Check the official overview at https://platform.claude.com/docs/en/managed-agents/overview for current options, since the product is in beta.
Does Anthropic store Managed Agents sessions?
Sessions are stateful and stored server-side, and you can delete them. Decide up front what data the agent may see and set a habit for deleting sessions you no longer need.