Claude Managed Agents Are A Game Changer (Here's Why)

Income stream surfers| 00:18:49|Apr 20, 2026
Chapters23
A preview of Harbor Build using the Claude agent SDK to automatically create a full website from user input, including a working example that can book via Stripe.

Claude Managed Agents unlocks building serverless, credential-safe agents that orchestrate tools to generate live websites in minutes.

Summary

Income stream surfers dives into how Claude’s managed agents and the agent SDK can run in a Linux sandbox to automate website creation and operations at scale. Theo walks through deploying a persistent, sandboxed environment that keeps state across long-running tasks, with MCP-backed credentials stored securely in vaults. He demonstrates Harbor Build, a forthcoming website builder that uses CLAUDE’s managed agents to scaffold an Astro project, preview via E2B, and connect to GitHub, Supabase, Vercel, and Stripe for live deployment. The video highlights the practical benefits: per-user vaults, 20-minute builds with no HTTP timeouts, and the ability to add bespoke tools like Bright Data scrapers on demand. Theo emphasizes that this approach dramatically reduces the complexity of orchestrating tools inside AI, enabling serverless, cloud-based workflows that were impractical before. He also contrasts it with traditional API usage, arguing that managed agents make it possible to grant scalable, secure agent access to client credentials without exposing sensitive data. The showcase includes live tool invocation (bash, read, write, edit, MCPs) and a quick code exercise to spin up a Bright Data scraper, underscoring how fast and flexible the system is. The takeaway is clear: with managed agents and Convex’s agent component, you can spawn your own Claude-powered automation hub anywhere, in seconds, and start offering powerful AI-driven services to clients. Finally, Theo teases Harbor Copilot and a CMS integration that will let users modify prices and products directly within Harbor, all powered by the same agent framework.

Key Takeaways

  • Managed agents run in a persistent Linux sandbox, keeping state across multi-turn conversations for hours-long work.
  • Credentials are stored in vaults and accessed via MCP, eliminating the need to hand out raw secrets to the AI.
  • You can add bespoke tools (e.g., Bright Data scrapers) directly into the agent’s toolkit without heavy reengineering.
  • Harbor Build scaffolds an Astro project, previews with E2B, and wires GitHub, Supabase, Vercel, and Stripe for live deployment.
  • The architecture supports serverless, cloud-native operation that can be offered as a service to clients, with secure credential handling and scalable tooling.

Who Is This For?

Ideal for developers and agencies looking to automate complex, credentialed workflows with AI. If you want to build or offer AI-driven website builders, scraping services, or CMS-enabled clients-facing apps with secure vault-backed credentials, this video is essential watching.

Notable Quotes

"This is a configurable agent harness that runs in managed infrastructure longunning tasks and async work."
Definition of what managed agents are and why they matter.
"We actually use managed agents which has MCP built into it and also vault scoped credentials which means we can get people's credentials, put them into a vault in anthropic system."
Credential security and MCP integration as a core benefit.
"With no extra coding whatsoever, we can add any tool you want, like Bright Data, to scrape websites."
Demonstrates tool extensibility inside the agent framework.
"This entire time Harbor Build has been running and has built me this website right here."
Live demonstration of a site built by the system.

Questions This Video Answers

  • how do Claude managed agents securely handle credentials in the vault
  • what is MCP and how does it integrate with Claude managed agents
  • can you add custom tools like Bright Data to Claude managed agents
  • how does Harbor Build scaffold and deploy a live site using AI
  • what are the benefits of serverless AI agents for client-facing services
Claude Managed AgentsAnthropic MCPBright DataHarbor BuildHarbor CopilotAstroVercelStripeVault credentialsLinux sandbox
Full Transcript
Okay guys, and look at this. This is a serverless environment running the clawed agent SDK. This was not possible until very, very recently. And in this video, I'm going to show you exactly how we're using the Claude agent SDK to do some pretty impressive things. This is a preview to Harbor Build, which is coming very, very soon. This will basically build an entire website for you based on the inputs that you give it. This is an example of one of the websites that it created for me as an example. And it works as in you can literally book things using Stripe. This is coming very very soon to Harbor, but that's not what this video is about. This video is actually about the Clawude agent SDK and managed agents. Let's jump into things. So basically, if you've used the Clawude API before, you've probably used the messages API, right? This is a very very basic conversational API where you can yeah have a conversation with AI. The managed agents API on the other hand is completely different and one of the main things it is is it's a configurable agent harness that runs in managed infrastructure longunning tasks and async work. That's the important thing here. Another really cool thing is you can add MCP servers, you can add skills and you can add a lot more. Okay, so this is what we're going to be talking about in today's video. So the agent architecture, how we build and deploy live websites with Anthropics managed agents API. serverless convex and a per user vault. So a user types a prompt, a site goes live. This is harbor build. So you prompt, user describes the site they want. We scaffold it. Agent builds an Astro project in a Linux sandbox, right? We sync to E2B which basically gives a preview like lovable, bolt, etc. And then once that website is done, they can claim the website or you will be able to claim the website using a process where you connect your superbase, you connect your GitHub, your versel, stripe and then you actually get the website at the end of it. We're on the beta managed agents API. So the important thing here is persistent sandbox, right? So bash, write, read across many turns. We don't replace state 3 sessions up to three-hour sessions. 20-minute builds don't hit HTTP action timeouts. Built-in MCP. This is the really cool thing. Instead of having to mess around with, you know, giving an interface whatever to the AI, we actually use managed agents which has MCP built into it and also vault scoped credentials which means we can get people's credentials, put them into a vault in anthropic system and that is where things start to get super super interesting. So yeah, basically this is the whole way that this works. The main thing is the Linux sandbox, right? Which is super super nice. Instead of using the messages API, we use the agents SDK inside a Linux standbox to basically you can create anything with this, right? And the really cool thing about this, by the way, guys, just to quickly show you, is it actually has these tools here built into it, right? So Claude manager agents gives Claude access to comprehensive set of built tools. bash, bile operations, read, write, edit, glob, etc. Web search and fetch and MCP servers. What does that mean? It means built into the system. So with no extra coding whatsoever, right? I managed to build a system that uses bash bash bash right bash bash bash. Right? Previously, this would have taken me so much effort to do without the managed agents SDK. Now, the really cool thing is you can add any tool that you want. It's not only these tools here. So, for example, and I've done this before with the a with the messages API, if you wanted to, for example, add something like bright data to your stack, which by the way, Bright Data, thank you for sponsoring this video. You can go and get yourselves $25 free credit with Bright Data with the link in the description. But like, all you would have to do is just go to like a piece of code that you want to use. So like let's go to playground here and let's select the unlocker API because that's the most useful one in my opinion. All you would do is either go to code examples or playground. Just grab this getting started right here. Right. Copy this. Give this to clawed code and tell it that you want a new tool that uses bright data in order to scrape a website. If you don't know the way bright data works is something like eyes suit, there's a huge problem with eyesuit which is that it's protected by cloudflare. So we actually cannot scrape uh eyesuit with normal scraping means which is why we use bright data to come along and scrape eyesuits. So without bright data we wouldn't be able to do this. So basically if you need something like a universal scraping system or if you need to scrape specific things like Tik Tok, Facebook, Google Maps, that kind of stuff, that's where bright data comes in. And we actually do use bright data inside Harbor. So you can see here, this is our usage of bright data over the last looks like month or so. We use it every single day to scrape people's websites that have Cloudflare. Not just that, we also use it as a backup scraper. So the way Harbor works is we scrape people's websites for images. If it returns no images, right, we then try a bright data scrape instead. And if that works, we take the images and we use them in the article. If it doesn't work from there, we try a few other things as well. Okay, so just so you guys know, this is the first um like I'm just showing you this as an example. This is what I've been working on for the last 3 days or so. Um, I'm really really excited to kind of release Harbor Build, but it's not quite ready yet, and we don't really know how to release it either. But it's literally one of the best website builders I've ever used. It's significantly better than Lovable. It's better than Bolt. It's better than, you know, whatever. It gives you SEO friendly websites ready to go as well. But that's not what we're talking about in this video, but I just wanted to show you what's actually possible with the ma with managed agents, right? Another thing as well that we're using managed agents for is actually Harbor Copilot which is coming more quickly. Again, I'm just showing you this as an example of what you can do with the agents SDK or with managed agents, right? So, this actually uses the same system. You can see it's thinking here and then um you know it basically tells you everything it can do for you. This is an agency account. So, it it tells you from the from an agency perspective what it can do for you. But basically again this just uses the same system where all of the things it can do are listed as tools inside the code. Right? So add a client if you say add a client it enacts the add a client tool and uses it to add a client to harbor. Right? So add a uh add a website to this account that is I don't know uh two men. Right? And you'll see it will invoke the tool here. There we go. You see find site map. That's tool number one. Create site tool number two. Got the site. Now creating the site. Okay. Blah blah blah etc etc. Right? These are all tools bespoke tools inside the API right inside our managed agent right. And this is so easy to code. It's ridiculous. It's so easy to code. So yeah guys, just go watch the video. There's a link to Bright Data in the description of the video and in the pin comment. If you need enterprise level scraping for whatever you're building, go and check out Bright Data. Let's jump back into the video. Okay, so this is kind of the magic here. This is the one agent that we build. This is the code that we actually use, right? So it creates an agent. It has all of the tools needed. It has all of the MCP servers needed and it just works because it's such an intelligent model. We actually just changed to Sonet 4.6 because Opus 4.7 uh yeah, a little bit too expensive to be honest with you. So that build that I just showed you here, this is actually Sonet 4.7, right? Uh sorry, 4.6. It's not Opus 4.7. So that's why I guess the build is not quite as good as an Opus build, for example. But it's it still looks pretty it still looks pretty good, to be honest with you. Yeah, you can choose the model, you can choose the system, you can system prompt, right? MCP servers, you can add bespoke tools, which is the important thing. If you need a tool to scrape, if you need a tool to add a website, if you need a tool to, you know, whatever, you can add as many tools as you want. That's the beautiful thing about this, right? And it starts the agent in a Linux sandbox, which you can find on Anthropics website. Okay, so these are all of the agents, right? So if you go to platform.lord.com, claude.com and you go to managed agents. These are all agents spawned from my code, but they actually also appear here. So, look, this is the really really beautiful thing, right? So, you have a system prompt. You could also use skills instead of system prompts. I thought I had skills configured, but I obviously didn't configure them properly. I could edit this and add skill, but that's not what I'm going to do. We have the MCPS, which is the super super interesting part of this, right? because you can connect to people's things without, you know, storing like they're stored in these credential vaults right here. So every single time you start a managed environment, a managed agent, it stores the credentials here so that you don't have to worry about storing these credentials, right? Which is super nice. So look guys, the entire time that I've been talking, this has been building in the background. I cannot wait to release this to people. It's not ready for production just yet, but when it is, holy crap, this is going to be amazing to show you guys. So yeah, if you think of this as like being able to spawn your own version of clawed code anywhere in the world at any time with people's credentials inside it, you can start to understand why this is so valuable because you can sell these to people, right? If you can say to a company, look, I can do all of the work that your employees do for you for much cheaper using AI, using this system, a lot of them are going to and not only that, but you will be in control of the system, right? I don't run it, you run it, right? That's when people start to get super super interested. So, yeah, like I said, the credentials thing is super important. you no longer have to worry about like how do I give access to um GitHub to an agent running in the cloud which was like a question that I couldn't answer in my own head but now it makes sense we store the um credential in a vault and then Claude accesses that vault when it uses an MCP right so super super nice super super interesting and then even if an MCP doesn't exist you can just use bash like bash scopes So versel and resend the agent has the ability in the cloud because it's running in a Linux container to run CLI commands right. So when it tries to connect to Versel which Harbor build will actually launch your website on Versel as well and also it puts like resend emails in there as well. So like if someone buys something from your Harbor built website, it will then send an email to those people and it does that not through MCP but through CLI because it's running in a Linux environment. And then this is just the launch session. So like how it actually works, right? This isn't that important. The initial user me message is where we inject. So this is how you create variance. So it doesn't just create the same thing every time. It's the user input, the prompt that you give it which changes the outcome of the entire build. And then the this is what we basically see here. So what the agent tells us, you can see here on the left, this is what the agent tells us, right? So we can see when it uses an MCP, we can see when it uses an agent message, we can see when there's an error, etc., right? Really, really useful for users, of course. And then we have four classes of tools. We have built-ins so bash, write, read, edit. We have mcps, github, superbase, and stripe. We have custom bash protocol and we have anti-tool. And that basically this anti-tool is from anthropic themselves trying to protect against prompt injection. And kind of this is the whole idea of harbor build right here. So we build, we patch, which is we make changes to people's websites, but the thread continues. It's the same conversation as before, right? So, three different types of agent all in one box, right? So, it's super super interesting stuff. And then this is what we would have to build without managed agents. You would need a way to have collord code in the CLI. So, you'd need to have like a VPS or something like that. 3 plus hour sessions. You'd have to somehow get MCP into it. The credentials thing is a nightmare. And also streaming is also a nightmare. But with managed agents and with Convex's agent component, you can build anything pretty much instantly using this system. Okay, so how do you actually build one of these agents as of today? What you what I would personally recommend is going to claude. Let's go new session here. I've been using the app more and more, guys. Cannot recommend it enough. So we'll call this managed agents. Then open. do a deep dive into managed agents and how they work, etc. I want you to be able to spawn a Sonnet 4.6 managed agent for me using the anthropic API. Start by researching it. So yeah, this is the first step. You need to understand how managed agents work. And there is actually a claude API skill that is now included in um claude code here. So I'm going to guess that that actually includes everything that I need here. So manage agents. B it's done already. Really really amazing stuff. Okay. So let's just build a very quick agent, right? So let's just say stop. Build me a quick working example in uh I don't know HTML. No, maybe not HTML. Yeah, HTML, CSS, JavaScript, and Python, which is a universal scraper that scrapes images using LLM scrape and bright data to scrape the content. It should extract JSON based on what the user inputs. For example, if they say, "Get me all the images from this uh page, it should output images in JSON." Ah, it's pretty cool that it can open this here here in the preview. I do absolutely love this from Claude Code. This is cooking. Theo for some reason on Twitter hates this thing. Someone told me that it was because they just replaced his entire SAS. But like this is this is crazy cool. Now obviously this is Oh, maybe I'll put it in light mode. No. So it's uh dark on dark, but that's fine. Okay. So there we go. Now let's just say two men. And let's say um extract pricing and product and product images. Scrape. There we go. So, unlocking two men via bright data. There we go. Beautiful. This is just insane that you can just do this this quickly. Look at that. Absolutely absurd, guys. Absolutely absurd. So, it made a custom tool which was the bright data scraper, right? And then I don't know why this isn't getting uh images and things. Not sure. I'd have to look into it in more detail, but this is just a very very quick example. Let's just try a different example. So let's do eyesuit uh which cannot be scraped by traditional means. It has to be scraped by something like bright data. And bright data is honestly preferred. The pricing of bright data is really really good as well. It's super cheap. And there we go. Look at that. So now we just created a universal scraper using managed agents in literally under 20 seconds. Right? This previously this was not possible. This is absolutely crazy. So, it has a custom tool which is unlocking with bright data and then a custom scraper and then a prompt that will do this for you. And you could do this all serverless, right? You could do this in the cloud. You can do this for people. It It's absolutely absurd, guys. This is replacing things like NAN for sure. You can connect it through MCP so you don't and uh vault so you don't even have to worry about storing people's credentials. Yeah. This entire time this uh Harbor Build has been running and has built me this website right here. Let's just see if Stripe works. It might not work. Yeah. So, cross- sight post form submissions are forbidden. That's fine. That's just um probably a cause issue or something like that. This is the website that was created by um Sonic 4.6. It's absolutely insane. Harbor Build. I cannot wait to show you guys Harbor Build. Now, the really cool thing about Harbor Build, which I really, really can't wait to show people, is you actually have a CMS built into it, right? So, if you want to change the price of something, if you want to add new products, um you can literally go and add new products using Harbor, right? And you can use Copilot as well by pressing Ctrl K to also help you add more products. All of this is only possible because of managed Asians and Opus 4.7. I've been cooking for over 40 hours since Opus 4.7 was released. So, for context, that was like four days ago. I'm absolutely knackered, guys. I'm going to leave the video there. Thank you so much for watching. Go and check out Bright Data. Go and check out Harbor SEO. And if you're watching all the way to the end of the video, you're an absolute legend. And I'll see you very, very soon with some more content. Peace out.

Get daily recaps from
Income stream surfers

AI-powered summaries delivered to your inbox. Save hours every week while staying fully informed.