Claude Managed Agents is Anthropic's hosted platform for building and running AI agents that complete real tasks — reading files, browsing the web, and calling the apps you already use — without you managing any infrastructure. Launched in public beta on April 8, 2026, it runs on Anthropic's cloud and bills at standard token rates plus $0.08 per session-hour of active runtime.
What are Claude Managed Agents?
Think of an agent as a narrowly focused digital worker assigned to a single job: it has its own instructions, its own tools, and access to the apps you use every day. Unlike a chatbot that only answers, a managed agent goes and does the work — deciding the steps on its own. What makes it "managed" is that Anthropic runs the whole thing on its own cloud. You don't build an agent loop, wire up tool execution, or stand up a server; Anthropic handles secure execution, state, permissions, and tracing. You can build one from the Claude Console, from Claude Code, or with a new command-line tool. Inside the Console there are four pieces to learn: the Agent (its instructions and tools), Sessions (a single task run), Environments (the secure cloud sandbox it runs in), and Credential Vaults (where your logins and keys are stored). You can start from a blank setup or a ready-made template like "Support agent" or "Deep researcher."
That consumption-based pricing matters for planning: you pay for the model usage plus the time the agent is actively running, so costs scale with how much work you hand it rather than a flat seat fee.
How is this different from a chatbot or a no-code workflow?
Tools like Zapier or n8n run a fixed, linear set of steps. A managed agent is different: you give it a goal and the tools, and it chooses the path — researching, calling APIs, writing files, and checking its own work along the way. And because it lives on Anthropic's infrastructure, you skip the part most people got stuck on: hosting and securing your own agents. Anthropic has been adding to it quickly. Recent updates brought cron scheduling, so agents can run on a recurring schedule, along with credential vaults for storing secrets. The security design is worth understanding: when an agent needs an API key, the real secret is swapped in at the network boundary — right before the request leaves the sandbox — so the model running the task never actually sees your key.
This is the release I'd tell every business owner to stop and look at. The barrier to agents was never the idea — it was the plumbing: code, sub-agents, servers, secrets. Managed Agents removes all of that and leaves you with the part that creates value: pointing a focused AI worker at a repetitive task you already do. My one caution: it bills to your API, not your Claude subscription. Set a budget, start with one tightly-scoped task, and improve it the way you'd refine any business process — adjust, test, repeat.
It's also a signal of where the market is heading. Early adopters already include companies like Notion, which is using the technology so teams can delegate work to Claude inside their workspace. When the tooling gets this accessible, the advantage shifts from "who can build an agent" to "who knows which tasks are worth handing one."
How do you build your first managed agent?
The Console walks you through a four-step wizard. First, Create agent: in Quickstart, describe what you want in plain English, or pick a template. For example — "Read my content calendar in Google Sheets, research competitor titles in my niche, and write a script plus five title ideas back into the sheet." Claude drafts the agent's name, system prompt, and tools for you to review. Second, Configure environment: choose or create the cloud sandbox it runs in, and set its access level. Third, Start session: store your app login in a Credential Vault, then launch a session, give it the task, and watch it work in the tool-use panel. Fourth, Integrate: once it runs well, deploy it on a schedule or call it from your own systems. Then keep refining the instructions and feeding it reference material until the output is exactly what you want.
Frequently Asked Questions
What are Claude Managed Agents?
A hosted platform from Anthropic for building and running AI agents that complete real tasks, while Anthropic manages the infrastructure: secure execution, state, permissions, and tracing.
Do Claude Managed Agents use my Claude subscription?
No. They run on your Claude API and bill at standard token rates plus $0.08 per session-hour of active runtime.
Do I need to code to build one?
No. You can describe the agent in plain English in the Claude Console and start from a template — no terminal required. You can also build via Claude Code or a CLI if you prefer.
How are my passwords and API keys protected?
Credentials are stored in vaults and substituted at the network boundary before requests leave the sandbox, so the model running the task never sees the actual key.
Can agents run automatically on a schedule?
Yes. Recent updates added cron scheduling, so agents can run on a recurring schedule, plus environment-variable credentials stored in vaults.
Claude Managed Agents takes the hardest part of AI agents — the infrastructure — off your plate, leaving a point-and-click way to put a focused AI worker on real tasks. For businesses and knowledge workers, it's the most accessible on-ramp to agents yet. Start with one small, repeatable job, mind your API budget, and refine from there.