I Tested Claude's New Managed Agents... What You Need To Know

Four days ago, Claude came out and said, Hey, you're no longer allowed to use your subscription inside of third-party harnesses like open claw. And people were upset because now they're paying a lot more for AI, unless they're using something local. And then yesterday Anthropic says, Hey, we literally have a model that's too dangerous to release. It's crushing Opus 4.6. So then today they come out with this thing called managed agents with the headline of getting to production 10 times faster, and honestly, I've spent the past three hours playing around with them and I'm a little bit disappointed. Because after they announced this and after they announced this, you think, okay, they're really coming for open call this time. This is basically just them saying, Hey, we're now allowing you to host an agent on our cloud and our own infrastructure, so you don't have to spend all that time actually setting all that up, shipping a production agent used to mean months of infrastructure work first, but now managed agents handle that for you, where you basically just define your agents, tasks, tools, and guardrails, and then they set up that sort of cloud sandbox environment for all of that to be ran in. And that's why they're basically saying build and deploy agents 10 times faster. Now they are actually doing some really cool things for certain businesses. We can see that notion has a managed agent set up where teams can delegate works to cloud directly inside of notion itself. They basically drag over some tasks to a different status and then cloud agent picks it up and processes them. And if you come to the docs, you can see a lot of other actual proof of what teams are building with managed agents. So in today's video, I'm going to do a quick breakdown on how they actually work, what they cost, when you should use them, and just kind of my honest thoughts and why it's a little bit disappointing to me. So if you head into the cloud console, you can see on this left-hand side that we have something called managed agents over here, and if I just click on quickstart, we can basically see a little interface where we have some templates to choose from, or we have a chat looking interface where we can just describe the type of agent that we want and it'll just build it. Now, the good news is you don't actually need a subscription. So those of you that might not have a cloud code subscription, all you have to do is come in here and just set up an API key, which you do have to put in like five bucks to get started. So let's say I want to create an agent that's able to analyze competitors and it will basically do research and then give me an analysis on what we can do differently as a business in order to stay ahead of the curve. So what happens is it automatically over here on this right-hand side, writes out that prompt. You can see, we have a name, a description, a model, a system prompt, and then we have other things like an MCP server and some tools. We also have the ability to give these agents skills, which is awesome. Of course. Now I could go ahead and create this agent or keep refining. So if I click on that, it'll say like, Hey, what else do you want to add here? What I'm doing is asking it to change the model to Opus 4.6 instead of Sonnet 4.6, and you can see it just changes that configuration right there. And now I will hit create this agent. So what I want you to think about as we're going through this is your perspective, if you're someone who's already living in cloud code, then this doesn't really seem that much to you and it's not super helpful, but if you are someone who is mainly living in cloud chat and you've never used coworker code, then this is a huge, huge value add for you. And you can build agents really, really fast by just describing a goal. And it's literally just a five-step process. You can see, we just finished step one, which was to create the actual agent. And now we have to create the environment, which is our cloud container that Anthropic hosts, and it has pre-installed packages. So we're basically just giving this agent its own little computer in the web. So it says right here, next, we have to configure the environment. Cool. We'll go ahead and click on next. An environment is like a cloud container, blah, blah, blah. It defines networking rules and basically says like, what can your agent actually do? So I have three environments set up, but I'm just going to go ahead and say, okay, let's create a new environment for this specific agent. So that's what I said. It says, what networking access does this agent need? I'm just going to go ahead for the sake of the demo and say unrestricted. And now it's creating that environment super quickly for us. Once again, we're not doing anything. All I have to do now is click on start session. So it says a session is live, but before we actually test, we need to connect to ClickUp. Now here's where it gets interesting. All you have to do to connect is. It's kind of like an OAuth login because this uses like MCP servers and things like that. So once again, the value prop here, if you've never built an agent and you're in cloud code or something, you might get lost when you start to think about like API keys or access. But here it's as simple as signing in and then it gets stored in your vault. So we have a vault that can be shared across different people in your team. I created one contest vault for the sake of the demo. Let's say I want to create a new vault in order to store this different ClickUp credential. It says that authorization is required to use this MCP. I could do an access token, or I could do OAuth credentials or right here, it just basically prompted me to do the actual like single sign on. And then we should be all set. I'll go ahead and click connect as you can see. And then all I have to do is choose which workspace I actually want to connect, so I choose up at AI. It's as simple as that. Basically anyone could be able to go through this flow and create an agent. And now it's already done. So it says that our session is live and we can go ahead and start chatting here. But the one thing I just realized is this has no idea about our actual business. So how could it give us a good report? So now I'm giving it a little bit more context about our actual business, and it's going to help us change the way that this agent is prompted. So if I switch over here to the agents tab, you can see that we have this agent, but you have to be really careful here because in the agents tab, we can see that I have two versions, right? Here's the system prompt for version one. And here's the system prompt for version two. It didn't even change. So I don't think that it actually implemented that knowledge. So what I would do here is I would come over and I would click guided edit. And then I could basically say, you need to add to the system prompt that we run a AI coding platform business. So our whole goal is to help people turn a natural language request into software or AI agents or websites, things like that. This is important so that you can actually give me a good analysis. And now you can see it was actually able to change the system prompt in here. So the agents outputs should be a lot better now. So let's go ahead and click test run. We have to choose the actual environment to run this thing in. So we're going to do competitor Intel, and now I could drop in a message like analyze cloud code, and it would actually go kick off and start this session up. And what we can do here is we can see step-by-step what's going on and how long each step is actually taking. If I click on view session, we get a bit of a better overview because we get like this horizontal bar, which kind of shows us the actual path you can see right now. It just did four different web searches. And it has basically all the same tools that cloud code does. So when you think of like your bashes or your reads or your greps or your searches or your fetches, that's basically what these agents are doing. This is essentially the, you know, cloud code, you know, the agent SDK, but it just has a nice wrapper around it, but it gives people who are scared to look at something like cloud code, a much nicer, regular Claude sort of interface. And we also have the ability to use this ask Claude button, which means that we could chat with Claude to understand how this is set up or what we're looking at or what we need to do to make changes. So anyways, while that's running, let's talk about what else we have over here. We just saw the quick start where we could use templates, or we could just use a conversation to start an agent. Obviously all of these templates are basically just prompts with maybe some servers connected that we would have to connect our credentials to. But then we have our agents. So here you can see that I've spun up four different ones. I've got the competitor intelligence one using Opus. I've got these other three using Sonnet and a few other that I've already deleted. But each agent basically has to work within an environment. And so that's why in this case, I basically have one environment for each of these agents. Now, what you get charged on is not these environments. You actually get charged on the sessions being active. So all of these are idle. They're not costing me anything, even though the environments are running. And what you get charged on is 8 cents per hour that these sessions are actually live and running, as well as any of the actual token costs that you're acquiring via the API. And then of course, down here, you have your actual credential vault where you could share these with different people. And within each of these vaults, you have different connections. Like here, you can see I did the OAuth to connect to ClickUp. If I click on add credential, we have all of these other examples that are already set up. And most of these are pretty standard where you just basically just get to sign in. You can see that this session is now idle. So if we click in, we can see what happened. You can see that this took about three minutes. It used this many tokens. And we can see kind of each individual step that the agent took here. So we have our actual competitive intelligence summary about cloud code. As you can see down here, it goes into things like scaling revenue, key strengths. But keep in mind, all of this system prompts was created by cloud itself. So we would probably have to get in there and make it a little bit more specific. Now, what you'll notice though, is this didn't even get sent to ClickUp, even though we actually had this connected. So that would be another case where I, the user, would have to go back and fix it there. So let me talk about some other things I tried. I made one called field monitor, which basically it looks at things that are going on in the tech space and then sends me a signal in my ClickUp channel. And this is what that output looked like. This one took me about two minutes to set up. I hit run and it's pretty good, right? I get all this information from the week. It gives me different clusters of information with all of these having sources, and it actually was able to send that to me in my ClickUp. And then I made another one right here called the ClickUp research agent, which the whole idea was that I would basically be able to come into this research queue area in my ClickUp. I would be able to create a new to-do down here for a different type of voice AI provider. And then when I move it to to-do, the manage agent would pick up the task and then it would do research. And it would leave me a comment over here. As you can see, I've got a summary, key findings, information like this. And then I get all the sources at the bottom and then it moves it to complete. Now, the thing about this is because our agents have to be triggered by a call to their API. It's just not super automatic. And it requires a lot of glue in order to actually put the infrastructure together, because what I had to do here is I had to come into the agent and say, Hey, go check the ClickUp and then see if you've got any to-dos and then fulfill them. And sometimes that's fine. But really what I wanted to do is I'd want it to pick up the to-dos automatically, which means I would need something from ClickUp to tell something to send an API call to my ClickUp research agent. And I was like, okay, cool. So if I can't do that via ClickUp webhooks, then maybe I'll just do a scheduled automation. So maybe I'll just have this ClickUp research agent wake up every 30 minutes, check the to-dos. If there's one, it'll process it. If there's not one, it'll just go back to sleep and that's fine. But it can't even do that. It can't even wake up on a cron. And so that's what made me realize that this managed agent thing, it really serves a purpose for a specific type of person. For me personally, I'm probably never going to use this because I can just do cloud code and make things better and more custom to my needs, because if I wanted this exact type of agent, which I've actually done before, and I've made a video on it, I would just use trigger.dev. And if you want to check out that video, just check it out up here, because trigger.dev lets me just host my code on the cloud and it can set up crons and things like that, and it can delegate to different scripts. So I would just way rather use trigger.dev than use this managed agent environment, and it's also a lot cheaper in trigger.dev, because I mean, you're telling me that Anthropic has a model that's so good that it cannot be released, but they can't just give us scheduled crons in the cloud? Because at this point I could literally glue it together by having a scheduled trigger in NNN, or having a ClickUp trigger in NNN, and then just making this shoot off an HTTP request to my managed agent over in Anthropic's cloud. But that's just like over-engineering. Why can't we just have a ClickUp trigger natively there, or a scheduled trigger natively there? I just... Now there are some other things that are coming, which I do think are going to make this a little bit cooler. There are three features that were teased, but all of these are not available yet. You have to basically go apply for early access. So there are going to be outcomes, which is defining specific success criteria for a task, and the agent will self-evaluate its work and iterate until it meets the criteria. So it's kind of like using Karpathy's auto-research, but built into the build phase, and I do think that that's pretty cool. There's also going to be multi-agent orchestration, where one agent can invoke other agents via a callable agents tool, and a coordinator agent delegates subtasks to specialized agents. And this is the multi-agent swarm pattern that is managed by Anthropic's infrastructure. And then what else I do think will be very cool is persistent memory. So memory that survives across sessions. Basically meaning, if I go to my sessions, let's say that, you know, we have these two that were ran by the competitor intelligence agent. Each time the competitor intelligence agent wakes up again, it is completely stateless. All it looks at is its system prompt. But what they're going to do is basically say, this agent will have memory across all of the sessions that it's running, and that is going to be pretty cool. Now, I do think there's a way that you could Frankenstein that as is, if you just give it the right things to look at and you have it write logs of what it's doing, but having that be a native feature will be pretty nice. But don't worry, I have already applied to this. I applied to this as soon as I saw the form multiple hours ago. So if I do get access, I will definitely be making a video about how it all works. The other thing that's interesting and we're talking about is the CLI because the CLI lets you build and use your managed agents right from cloud code. So maybe there's also some value there because in some cases, what you could do is not only build a managed agent in cloud code in a project that already has like a ton of information about you. Um, but you could also then build front ends for it because it already understands the configuration and what it needs, and it also means if you wanted to delegate some stuff to sub agents, you could do that from cloud code and not store them here. You know, sometimes if you have like a dot cloud and you have an agent's folder, that's going to eat up some context and maybe they don't get invoked in the right way. But if you have each agent pointed at a specific API endpoints that you could then call from cloud code, maybe there is some value there. But anyways, what I did in this project is I basically set up a managed agent in managed agents. So let's say I wanted to build this automation, right? And I'm in cloud code. So I just shoot this off and I'm telling it to use Anthropic CLI to build myself a managed agent. The goal of this would be, I can drop in a YouTube video transcript and then I can just have it do things for me in my ClickUp, like sending me a summary, as well as potentially some action items from that video. If I found out, Hey, you know, I want to implement some of these strategies into my workflow or whatever. That's cool because I could drop in the transcript here. It could send off that API call to my managed agent in Anthropics cloud. And then over there, it would take care of all of that ClickUp stuff. And I could just keep coding in cloud code right here. And like I said, it can also use all the context. So let's say I've got a ton of information right now about my business, about my quarterly goals, about our current projects. It would be able to basically use that in the system prompt while creating this managed agent so that the actual managed agent is a lot more specific and tailored towards us. So that's where I see a lot of this value with the CLI. Now, one thing you do have to be very careful about if you're building in cloud code is the credentials, because it might try to default and send over potentially in the system prompt, the API keys for the certain endpoints that you need, rather than if you're building it inside of this environment over here, it's basically going to default to using MCP servers. So just keep that in mind and be careful when you're doing stuff like that, because other people might be able to see those keys. And of course, if it's being hosted by Anthropic and in the conversation history, then it's there too. But really, if you're curious what I would do, and basically the way that I teach myself everything is I create a cloud code project for it. It created me this master guide on what cloud managed agents are, how they work. And I just had it do research on its own so that it could explain things to me and help me figure out how to use it. And then just say, Hey, you know, like I've also been using trigger.dev. I've also been using, you know, the desktop scheduled automations and cloud code. What's actually the best tech stack for me based on me wanting to achieve XYZ outcome, because it's always not a matter of which tool is the best. It's which tool is the best for this specific use case. And I think that that's just the healthiest mindset to have around building automations and agents. And now you can see, we just had this agent created in our managed agent dashboard, which is the YouTube transcript analyzer. I can click in here and I could see the actual system prompt that was created, which you'll notice is much more robust than the ones that the quickstart version had us build. So a couple of things before we wrap up here, when you're comparing this managed agent to the agent SDK plus trigger.dev or something like open claw, that's where you'll notice there's so many differences. And I'm not going to read this whole thing, but if you want to pause it and give it a look, feel free. But like I said, there's really two things about open claw that I love. And the reason why I still use it personally, and that is the heartbeats and that's the ability to talk to it from telegram so easy. And I think that that's what Anthropic really needs to be doing with cloud code that would just make it an absolute powerhouse. Heartbeats just basically means cron every 30 minutes, every five minutes, you can have open claw, just wake up and do things. And that's basically why it feels like it's an always on assistant. Besides the fact that every time you shoot at a message, it automatically wakes up as well. Now, of course you can do schedule tasks in the desktop app of cloud code, which I do, and I have tons of them up and running and I've got more on my other account as well. But it's just not the same as being able to do a heartbeat check every single five minutes to just check in on something. And that's what I said earlier. At this point, I'd rather use cloud code to build me something with the agent SDK, which is very similar. And then I would just say, Hey, trigger.dev run this every five minutes. So the who should use what, if you're just getting started with AI agents, this is where you might want to try out managed agents so much easier. No infrastructure to worry about hardly any API keys to worry about. But if you've already been building, then there might not be as much value as you might think. But anyways, this whole document that we've been looking at here, it, it does some pretty in-depth research and comparisons about the actual value of managed agents. And I will put that in my free school community for completely free. The link for this will be down in the description. All you'll have to do is find the post associated with this video and there will be the full PDF right in there. But anyways, I hope you guys enjoyed the video. If you learned something new, then please give it a like it helps me out a ton. And as always, I appreciate you guys making it to the end of the video and I'll see you in the next one. Thanks everyone.