New Claude AI Agent Update is Awesome!

追加: 2026-05-23

[00:00:00]New Claude AI agent update is awesome.

[00:00:02]What if the biggest Claude update yet just dropped and you missed it? Why are top builders already running agents inside their own systems? Is your AI workflow about to feel completely outdated? There is a brand new Claude agent update. Changes how you can use Claude at work forever, and almost nobody's talking about it yet. Hey, I'm the digital avatar of Julian Goldie.

[00:00:20]People learn and actually use AI tools so they save time, build better workflows, and grow faster with AI. This video I'm breaking down the new Claude managed agents update Anthropic just announced live from Code with Claude London. I'll show you what it is, why it matters, and how you can start using it today. Stick around because the last part is where this gets really interesting. Let me start with the basics. Claude is the AI system built by Anthropic. Writes, codes, plans, and now it runs full agents that can take actions for you. People use Claude through the chat app, through the API, and through Claude Code in the terminal.

[00:00:49]It's one of the most trusted AI tools out there for serious work. Now, here is the new thing. Anthropic just announced two huge updates inside something called Claude managed agents. The first one is called self-hosted sandboxes. It's in public beta right now. The second one is called MCP tunnels. That's in research preview. You can request access for that one. So, what does this actually mean?

[00:01:07]Plain English, Claude can now run agents inside your own systems. The agent loop stays on Anthropic's side. That part handles the planning, the memory, the error recovery. The actual tool execution moves to your environment.

[00:01:18]Your files stay put. Your network policies still apply. Your audit logging still works. It just plugs in. This is massive. Before this update, if you wanted an agent to do real work for you, your files had to leave your network.

[00:01:28]Your data moved around. For a lot of companies, that was a deal-breaker. They wanted to use Claude, but they couldn't let sensitive stuff leave their walls.

[00:01:35]Now, that problem is solved. Imagine you run a brand new lead magnet campaign for the AI Profit Boardroom, and you want an agent to handle the whole content production system without any of your assets leaving your own setup. You could spin up a sandbox, point Claude at your brand guidelines and past hooks, and let it draft an entire series of social posts that pull people into the AI Profit Boardroom. That kind of safe, controlled workflow is exactly how serious operators build at scale. You also get to choose your sandbox client.

[00:02:00]That means you pick the tool that runs the actual agent environment. Anthropic announced four supported providers right out of the gate: Cloudflare, Daytona, Modal, and Vercel. Each one does something different. Cloudflare runs sandboxes using micro VMs and lightweight isolates. Gives you control over outbound traffic. You can route, audit, or modify what leaves the sandbox. You also get zero trust secrets injection. That means credentials never get exposed to the agent. They get injected at the network edge. Company called Amplitude is using this setup to build a design agent for on-brand mockups. Daytona is different. Runs full composable computers, long-running, stateful. The same setup can handle a quick job or run for hours. You can connect to it over SSH. You can pause it and restore it later with full state.

[00:02:42]Clay is using Daytona to run an agent called Sculptor that builds and test workflows on its own. Modal is built for AI workloads, sub-second startup, scales to hundreds of thousands of concurrent sandboxes. You get CPU and GPU on demand. If you need to run heavy compute, this one is built for that.

[00:02:57]Vercel sandboxes combine VM security with VPC peering and millisecond startup. The Vercel sandbox firewall injects credentials at the network boundary, so they never enter the sandbox itself. A platform called Roggo is using this setup to build an analyst agent for finance work. And the cool part is you don't have to pick one. You can bring any sandbox client you want.

[00:03:15]These four are just the ones that work out of the box. Now, let me talk about the second big update. MCP Tunnels. MCP stands for Model Context Protocol. It's basically how Claude connects to outside tools. Think of it like a plug. You plug your tools into Claude through MCP, and now Claude can use them. Databases, internal APIs, knowledge bases, ticketing systems. All of that becomes something an agent can call. Before MCP Tunnels, if you wanted Claude to reach an internal tool, you had to expose that tool somehow. Maybe a public endpoint, maybe a workaround. The way it added risk. With MCP Tunnels, you deploy a lightweight gateway inside your network.

[00:03:49]That gateway makes one outbound connection. No inbound firewall rules, no public endpoints. Traffic is encrypted end-to-end. Your agents can reach internal stuff without you opening any doors to the outside world. Now, here is something you really want to hear if you've been following Claude updates. Inside the I Profit Boardroom right now, we go deep on this exact topic. There are full walkthroughs on setting up Claude agents the right way.

[00:04:09]There are coaching calls where you can bring your actual project and get help live. There are road maps built around using Claude managed agents for content systems, lead generation, and product workflows. There are prompt libraries you can plug straight into your own sandbox setup. Everything you need to actually use this new update is in there, organized by skill level. If you're watching this and thinking, "Okay, I want to actually do this."

[00:04:28]That's where you go. So, who actually needs this update? If you work in a company that handles sensitive data, this is for you. Now, you can run Claude agents and keep everything inside your own walls. Compliance teams, stop worrying. Security teams, stop blocking you. If you build with AI for clients, this is also huge. You can offer to run agents inside the client's own infrastructure. That removes one of the biggest objections when buying AI services. They don't have to trust a third party with their data. And if you're a developer building agent products, you don't have to hand-roll the sandbox layer anymore. You don't have to build the security wrapper.

[00:04:57]Claude managed agents gives you all of that, so you can focus on the actual product. Even if you're a solo creator, this matters. You probably have stuff you don't want sitting on someone else's server. Client briefs, personal notes, half-finished projects. Now, you can keep all of that inside your own setup and still have a full agent working on it. That used to be something only big companies could pull off. Now, one person with a laptop can do it. It also means your prompts, your workflows, and your private playbooks, they're yours.

[00:05:20]Thing leaks out, nothing gets logged somewhere you can't see. You own the whole pipeline end-to-end. Here are some quick tips on how to get the most out of this update. Tip one, start with the docs. Anthropic put out a full docs page on self-hosted sandboxes. They also have cookbooks on GitHub that walk you through setting up each provider step-by-step. If you're new to this, start there. Tip two, pick the right sandbox provider for your job. If you need fast cold starts and millisecond response times, look at Vercel or Modal.

[00:05:45]If you need long-running stateful sessions, look at Daytona. If you need fine-grained network control and audit, look at Cloudflare. Match it to what you're actually building. Tip three, use MCP tunnels for anything internal. If you have a knowledge base and internal API or a database that needs to stay private, set up an MCP tunnel for it.

[00:06:03]Don't expose your internal tools to the public internet just to use them with Claude. Tip four, pay attention to resource sizing. Self-hosted sandboxes, you control the compute. If your agent is running long builds, image generation, or anything compute heavy, give it the CPU and memory it actually needs. Don't starve it. Tip five, think in terms of agent workflows, not single prompts. This update is built for agents that do real work over time. Plan out what you want the agent to do. Break it into steps. Let the sandbox handle the execution while Claude handles the thinking. Here's an example of what's possible. You could use this setup to build a full content engine that creates videos, captions, and email sequences for the AI Profit Boardroom from one workflow. The agent runs inside your own sandbox, pulls from your private notes through an MCP tunnel, and pushes out content on a schedule. That kind of system would normally take a team to build, and now one person can run it.

[00:06:50]So, this is a really big update. Tropic is making it possible to use Claude agents in environments that were locked off before. Healthcare, finance, government. All of these spaces have strict rules about where data can go.

[00:07:02]Now, Claude fits into all of them. It also shows where the whole AI agent space is heading. The future is not about one giant cloud running everything. It's about agents that can run anywhere, inside your company, inside your home network, inside a customer setup, with proper security baked in by default. And the part most people are going to miss is how this changes what one person can build alone.

[00:07:21]You don't need a security team. You don't need a DevOps team. You can spin up a sandbox on Cloudflare, Daytona, Modal, or Vercel and have a working agent set up in a day. That is wild compared to where we were two years ago.

[00:07:32]If you want the full process, SOPs, and 100-plus AI use cases like this one, join the AI Success Lab. Link's in the comments and description. You'll get all the video notes from there, plus access to our community of 58,000 members who are crushing it with AI. And here's the final thing. If you're about to go try Claude managed agents for yourself, you're going to hit a few walls. Which sandbox provider should I actually pick?

[00:07:52]How do I set up an MCP tunnel? What do I do when the agent gets stuck? That's exactly what we work through inside the AI Profit Boardroom. Live coaching calls where you bring your setup and we walk through it together. Full tutorials on each sandbox provider. Roadmap built around Claude agents that shows you what to build first, second, and third.

[00:08:08]Prompt libraries ready to drop into your own sandbox. If you want to actually use this update and not just watch a video about it, come join us. Head to AIprofitboardroom.com.