I Built a Claude Managed Agent in 30 Minutes. Here's How They Work and Why They Matter.
Claude managed agents are the infrastructure behind AI-powered products. Here's what they are, how they work, 5 real use cases, and a full build demo.
I had a completely different article planned for today. It was almost done. I was going to polish it this morning and publish it in a few hours. Then I checked the news, and saw that Anthropic had just released Claude Managed Agents.
I couldn’t help myself. The other article can wait another week.
I even dropped this into our community chat:
And then Mia Kiraki 🎭 dropped this, which might be the most accurate meme about my life right now:
At this point, building with Claude feels like trying to document a moving train while riding it. I wrote about skills. Then about writing skills. Then Cowork. Then plugins. Then scheduled tasks. Then Claude Code. Then computer use. Then the AI Blew My Mind MCP and a bunch of other skills.
I don't seem to be able to stop writing about Claude.
And with Anthropic heading toward IPO, this is not accidental. They keep releasing features that get the entire community talking. It works because the products are good enough that we can’t help ourselves. I am, for sure, guilty of that.
So instead of the article I had planned on Perplexity Computer, I spent today testing Managed Agents.
I went through the case studies, the API docs, and what companies are already building with it. Then I built one myself, a feedback agent for AI Blew My Mind, connected it to Notion, and deployed it to a live website using Claude Code.
This is that article. Everything I found, everything I built, written in real time.
Here’s what we’ll cover:
Let’s get into it.
But first… what is an AI agent?
To understand why Claude managed agents is a big deal, I need to first clear up a confusion I had myself and that you might have too.
When you hear “AI agent”, you probably picture something like OpenClaw, Claude Code or Cowork. You give it a task, it figures out what needs to happen, and it does it on its own, autonomously. Researching, writing code, fixing errors, calling tools, checking its own work.
But those aren’t single agents. They’re orchestrator agents. Systems that manage multiple smaller agents behind the scenes.
Here’s what I mean. When you ask Claude Code to build something, it doesn’t do everything itself. It breaks your request into pieces and hands each piece to a specialized agent. One agent handles file editing. Another runs your code. Another catches errors. The orchestrator coordinates all of them, passes information between them, and keeps the whole thing moving until the job is done.
You've already been using these smaller agents without realizing it. Every time you use Claude Code, that's agents working behind the scenes. When Cowork organizes your documents and runs scheduled tasks, same thing. The skills and plugins behind the AI Blew My Mind MCP? Also agents.
Each one of those smaller agents is pretty straightforward on its own: it has a set of instructions, a brain (the AI model), and tools it can use. It receives a task, reasons about what to do, calls the tools it needs, checks the results, and loops until the job is done.
Keep this in mind, because that's exactly what a managed agent is.
So what’s new with Claude Managed Agents?
Not the concept. We've been building with agents for months. What's new is where they run and who they're for.
Claude Code and Cowork run agents on your computer, for you. They’re personal tools. Brilliant ones. I use them every day. But they serve one person at a time.
Managed Agents is for people who build products and platforms.
If you have a SaaS tool your customers use, this is for you.
If your company has internal systems your team works in every day, this is also for you.
Anywhere you want AI agents serving multiple people at the same time, on their own infrastructure, that’s what managed agents are built for.
What are Claude managed agents (and who are they for)?
Now that you know what an AI agent is, let’s look at what Anthropic just released and why it matters.
Claude managed agents are not a product. They’re infrastructure.
Claude Code and Cowork are products you use. You open Claude Code, it helps you write code on your laptop. You open Cowork, it helps you with files and tasks on your desktop. They’re personal AI tools. Powerful, but personal. They run on your machine, for you.
Managed Agents is different. It’s not a product. It’s infrastructure you build on. It’s what other products are made of.
Think about the difference between using Google Docs and building Google Docs.
When you use Google Docs, you open it, type your document, and close it. That's Claude Code and Cowork. When you build Google Docs, you need servers that handle millions of users, each user's files kept separate, and the whole thing running even when no one's at the office.
That's the kind of problem Managed Agents solves.
Before this, if a company wanted to add AI agents to their product, their developers had to build all of that from scratch: servers, prompt management, file handling, data isolation, scaling logic. Months of engineering work.
With Managed Agents, the developer defines the agent and Anthropic runs it.
How managed agents work
There are four building blocks:
Agent is the definition. The system prompt, the model (Sonnet, Opus, or Haiku), and the tools it can use. A customer support agent and a research agent use the same infrastructure. The difference is the instructions and the tools.
Environment is the workspace. A cloud sandbox where the agent does its work. You configure what’s installed, what it can access, and what files it starts with. Each user gets their own isolated sandbox.
Session is one run. A user sends a message, the agent works until it’s done. Sessions are persistent, so you can check progress or send new instructions mid-task.
Events are the log. Everything between the user and the agent, every tool call, every decision, every output, tracked so you can see exactly what happened.
The built-in tools include bash commands, file operations, web search, web browsing, and connections to external services through MCP servers. And there are research preview features already in testing: multi-agent coordination, memory across sessions, and self-evaluation.
Who should use managed agents
I think it’s important to set the right expectations.
If you have a product, this is for you. An app, a platform, an internal dashboard, a system that connects your own tools, anything where people interact with something you’ve built. It doesn’t have to be a big company. Even if you’re a solopreneur who built something that stitches your own systems together into one workflow, like Kim Doyal did with her 33-agent AI operating system, this is the infrastructure that makes it scalable.
If you’re a developer, you get the most immediate value. This replaces weeks of backend work. No sandboxing, no session management, no user isolation logic. You focus on what the agent should do, Anthropic handles the infrastructure. Ship in days instead of months.
If you're neither of those and don't have a product, this isn't for you yet. But if one day you start building tools for your own needs or for others to use, that's when Managed Agents becomes relevant.
How much does it cost
The pricing has two parts:
Token costs are the same as the regular Claude API. Whatever model you pick, you pay the standard rate per million tokens. For Sonnet 4.6, that’s $3 per million input tokens and $15 per million output tokens. Haiku is cheaper, Opus is more expensive.
Runtime fee is $0.08 per session-hour. That’s the cost of keeping the cloud container running while your agent works. Idle time doesn’t count.
Now, let’s put this in perspective: a customer support agent processing a 20-minute ticket costs roughly $0.03 in runtime plus $0.10 to $0.50 in tokens. That’s under a dollar for a task that might take a human 30 to 45 minutes. A continuously running agent would cost roughly $100 to $200 per month depending on usage. For comparison, a virtual assistant costs $500 to $2,000 per month.
What companies are already building with managed agents (5 use cases)
These are companies that had early access, built managed agents into their products, and shared their results.
Notion built a system where teams can assign work to AI agents from inside their workspace. Multiple agents run in parallel from a single task board, handling everything from shipping code to building websites and presentations. They saw a 90% cost reduction and 85% latency improvement through prompt caching. This is where project management is heading: agents that don’t just organize your tasks but do some of them.
Rakuten has 70+ businesses, and they deployed specialist agents across product, sales, marketing, and finance. Each agent took about a week to ship. Their release cycle went from quarterly to every two weeks, and critical errors dropped by 97%. The part that stuck with me: their team talks about delegating goals instead of tasks. Instead of “write this code”, they say “make this dashboard work” and let the agent figure it out.
Asana created what they call “AI Teammates”, agents that get their own accounts inside projects with permissions and everything. They built a Campaign Brief Writer, a Launch Planner, a Compliance Reviewer. What used to take days now takes a 15-minute review. And you don’t need engineering skills to set one up.
Sentry already had an AI tool that could tell developers what caused a bug. But developers kept asking: if you know what’s wrong, why can’t you just fix it? So they built that. Now the agent analyzes the problem, writes the fix, and opens a pull request. A single engineer shipped the integration in weeks.
Vibecode built a mobile app builder where you design, build, and publish production-ready apps from your phone by talking to Claude. What used to cost $10,000-$50,000 now costs about $100 and takes under an hour. This is the clearest example of what Managed Agents enables for me and you. The complexity is hidden. You just describe what you want.
Building a Claude managed agent from scratch (with demo)
I wanted to do more than just explain how Managed Agents work. I wanted to build one and show you the whole process, from the Claude Console to a live landing page.
So I built a feedback collection agent for AI Blew My Mind and connected it to a live website. Here’s what it does, how I built it, and everything that happened along the way.
The idea: A landing page where you can share feedback on anything AI Blew My Mind: the newsletter, the MCP, AI Migrator, the LAB, MyGhosts. Behind it, a managed agent that knows all of my products and has a real conversation with you. It asks clarifying questions, suggests existing solutions when relevant, and digs deeper into what you mean. When the feedback is clear, it saves everything to Notion.
Why a managed agent and not a form? Because a form just collects text. An agent can have a back-and-forth, clarify what you mean, point you to features you didn’t know existed, and turn vague feedback into something I can act on.
Everything starts in the Claude Console.
Below is the full walkthrough: every prompt I used, every step in the Claude Console, the videos, the screenshots. If you want to see how I built this from scratch, upgrade to a premium subscription.
This part is for paid subscribers. With a premium subscription, you get:
Weekly paid editions of the newsletter (like this one)
The MCP and all my premium prompts and workflows inside it
Full access to the AI Blew My Mind LAB: documented builds, automation chains, step-by-step guides
50% off AI Migrator, and $100 in credits for MyGhosts app