Claude Managed Agents: what they are and when your organization should use them
In brief
Managed Agents run Claude on long, multi-step jobs in Anthropic's cloud, so you don't build the infrastructure. You pay normal token rates plus $0.08 per session-hour while an agent runs. Still in beta. Here's who it suits, what it costs and how to start.
Contents
Short answer: Managed Agents let Claude run long, multi-step jobs in Anthropic's cloud: it reads sources, uses tools, runs code and hands back a result, while Anthropic handles the servers, sandboxes and sessions. You bring the instructions and skills. You pay normal API token rates plus $0.08 per session-hour of active running. It's in beta on the Claude Platform (the API), not in the claude.ai app.
Sources: Anthropic's launch post (April 8, 2026) and Managed Agents docs.
What do you supply, and what does Anthropic run?
| You supply | Anthropic runs |
|---|---|
| Instructions, skills and task logic | Sessions and step-by-step coordination |
| Which tools and connectors it can use | Tool execution in an isolated sandbox |
| Budgets and permission rules | The infrastructure, scaling and logs |
Who is it for?
- Good fit: teams that want a repeatable job (research, reports, data clean-up, code changes) to run on its own, without building and hosting an agent system.
- Maybe not: teams already running their own agents on the Agent SDK. At high volume, your own infrastructure can be cheaper.
- Not needed: one-off or personal tasks. In claude.ai, Claude already does long, multi-step work in a normal conversation. See Cowork and chat are now one Claude.
What does it cost?
- Tokens at the normal API price for the model you choose. See which model to use.
- $0.08 per session-hour, only while the agent is actually running.
- Tip: test locally first. Building and testing before you deploy costs only tokens.
- Set limits before you scale. See budgets and guardrails.
Good first jobs take more than 30 minutes by hand, pull from several sources and end in a document, like a weekly competitor summary or a client assessment.
How do you start?
- No code: the agent builder in the Claude Console lets you set up an agent and test it with a built-in session runner.
- From files: install Anthropic's
antCLI, sign in withant auth login, then runant apply <agent-file>. More in ant apply: resources as code. - Control what it can do with permission policies.
Can we run it on our own infrastructure?
Yes, in two ways, per Anthropic's docs: run just the sandboxes on your infrastructure, or run the whole service in your own cloud or data center. The full self-hosted version (November 12, 2026) is a private beta; ask your Anthropic account team.
Quick answers
How much do Claude Managed Agents cost? Normal API token rates plus $0.08 per session-hour while the agent is running. Idle time isn't billed.
Are Managed Agents part of Claude Team or Enterprise? No. They're a Claude Platform (API) product, billed through the Console, separate from claude.ai seats.
Is Claude Managed Agents generally available? Not yet. It's in beta, and some parts, like multi-agent coordination, are in research preview.