Snap Your Fingers and it's Done - Manus AI Agent

Welcome back everybody to another video. Today we're going to be talking about the Manus AI agent, and earlier this week it kind of blew up in the AI niche because it's very capable. It has its own Linux sandbox that it can operate in and move around and edit files. It can also browse the web and download things just like a person can. It's probably the most capable AI agent we've ever seen to date. But there are absolutely some caveats to this at the moment. As you can see from this post, 2 million people are currently on the Manus AI wait list, which is a lot. I doubt that the Manus servers or any other servers that Manus uses could handle that many people using the platform at once. So they're kind of in this slow rollout, this beta phase right now, and only certain people with access can use it at the moment. Thankfully, Manus did reach out to me and get me an access code, so I will be showing you some examples of my own in today's video. But even with the access code I ran into quite a few rate limits. This thing absolutely eats up AI computing for sure. Now another caveat to note is that Manus is also kind of a Claude wrapper. It's not its own in House LLM. It uses the Claude 3.7 Sonnet API, which is great at coding and great for a gentic workflow such as this and the actual computer use. You know, the ability to control the mouse, browse the Internet, change files inside of a Linux machine is using an open source agentic workflow. But again, they did do a lot of their own modifications and building on top of these platforms that they started with. So it's its own thing for sure. But make no mistake about it, it's not like all of this is from the ground up Manus built. Ethan Mollick here goes over some pretty nice key details I want to highlight before we dive into our own testing. He notes it's a Claude rapper, but a very clever run. It still runs into some of the same issues as general agents, including getting stuck every once in a while, which I also experienced. But it's also capable of some very good stuff, which I also agree. I had some pretty cool little demos that blew my mind and said I've never seen an AI agent do that before. Here's an example where Ethan had it get 10k for Apple and then visualize it in different ways to show me various trends and details. So this is a data analysis prompt, comes out here doing pretty good making an 8 page little PDF where you can go through revenue and net income with different graphs. Margin analysis year over year, growth rates from 21 to 2024, estimated revenue composition. Pretty darn cool. It's very much a capable agent. Even provides us with a nice TLDR here. If you've used other agentic systems like Claude Code or OpenAI's Deep Research, you'll have a good sense of what this can do and what the limits are likely to be for those who haven't used them. I suspect a lot of people will be surprised at what LLMs can really do. Totally agree, but if you're a viewer of this channel, you've seen us mess around with things like Deep Research, so you know that these agentic capabilities are something we've been looking into for a while and they have matured a little bit but aren't all the way there. So here is the Manus interface you log in with Google. As you might expect, there's not too much going on here. We've got a typical chat GPT esque interface, all your different sessions here on the side that you can close out or open. It actually has its own knowledge base. So if you're doing a larger project like using it to help you develop a full scale video game or something like that, you could probably add all of that knowledge you need it to understand in here. And and it can basically just be in the form of text which you could just kind of drop files in here and then have them automatically download into the Linux interface. But I don't think we're quite there yet. Also got some pretty simple settings here, language, appearance, profile, stuff like that. Now here is the first prompt I tested out with Manus. Pretty simple agentic ask help me find the cheapest Buy it Now version of a specific camera on the web. So obviously I'm looking to get a discount by buying used. I want a Buy it now option that is the cheapest one it can find and this is a pretty specific camera as well. You know this isn't some Canon that's super well known. It's a very niche and I thought that would be a good test. So it helps me by searching across the web here various Google searches like Fujifilm XE4 buy it now price, Buy it now ebay lowest price Clicked on a couple of ebay links, continued to search across a couple of other marketplaces as well. Amazon B and H photo and Video Adorama all sites that I've used in the past for camera gear. So Pretty happy with that. Once it completed the research, this is what it gave the cheapest being a silver body on ebay for only a thousand bucks, which is pretty cool. Then some from Keyh cameras ranging from 1160 to 1340 bucks, and then some new ones on Amazon with a kit lens for 2,000 bucks. And then some other used options for pretty expensive, a little under two grand on Amazon. Oh, actually, my bad. I don't think I mentioned it, but the cheapest option is indeed one from Adorama here for under a thousand bucks, which is pretty cool. Normally, you know, this would take me like 20 minutes to search and find the right one, but Manus can do it in just a couple of minutes automatically for me. Now another really interesting thing that you can do with Manus is you can actually get a little animation of it working. So we'll get the replay of this session public access, and if we go ahead and open this in a new window, you can see this is a sped up process of it searching through the web and finding all the different bits and pieces to get me the cheapest listings possible. When it's actually working. It doesn't move this quick like you're watching it right now. This is definitely a lot more sped up, but it's quicker than I could probably do it myself, that's for sure. Now we will move on to a task that it failed. This was an interactive stock market charting app with various insights. I wanted it to go through the general stock market over the past 365 days with notable insights that I can hover my mouse over and see what affected the stock market at that time. So not a super complex idea or anything. And it said, okay, you know, I'll help you create this. And by the way, this is also in a high effort mode which Manus seems to have removed at this point. It was an experimental mode where it would try to really complete the task for you, but as you can see, it clearly didn't end up working. In this scenario, as we'll go over, it sets up the application project structure, necessary dependencies get installed. It's going to use React with recharts for visualization. Yahoo Finance API is what it's going to use for the stock market stuff. It goes through installing dependencies, getting data from the last 365 days, notable market events, and then it encounters an API rate limiting issue when trying to fetch real stock data. So it decides to create mock stock data, which is really not that useful for me. At the end of the day that's not what I would want it to do. I wish it would ask before doing something like that. But anyways, it got stuck. Anyways, it's still thinking 50 hours later. You can see down here, it's still thinks that it's working. I don't think that it's using anything. I think it's just been cut off entirely from the system. But you know, in the app on the website here it still appears and is counting up like it's still thinking and working on this task over 50 hours later. Obviously I think it's done and it didn't work clearly. But yeah, here's an example of an unfortunate failure. It got stuck, was unable to complete this task even in that experimental high effort mode. So scrapped that one, moved on to building an $1,000 budget gaming PC using PC part picker. I wanted the best gaming for the buck. So let's check out the replay for this one. As you can see, it goes over starts researching different components, searches all over online to find out what's the best at the very moment for a thousand dollar gaming rig. And again you're seeing this in a sped up fashion. But it does it in, I'd say, you know, under 10 minutes at the very least. It goes through quite a few websites, it makes a plan for itself like an actual to do list, which is pretty cool. And then we ended up with another error. High system load has caused an internal server error. Oh yeah, you can see why it's only on a wait list right now. If they open this up to the public, it would be a disaster of drastic proportions. Considering it's even failing in this beta, this is definitely not a very difficult task. I could ask Chat GPT with basic web search to do this. So I'm a little disappointed that Manus failed it. You're like let's toss this into GPT4OMNI with search real quick. It looks like it actually went to PC Part Picker and just found a pre built made by someone else. Hey, you know what? It's a solution that works. We got a Ryzen 5 5600X. I mean I think that this has been the best budget gaming CPU for a long time now. Typical MSI board DDR4 at 3200 to 3060. Yeah, seems like a pretty bog standard thousand dollar gaming PC build though. Maybe Mandus is a little overkill. Even though it failed this task, this was probably my favorite example that I ran across. I don't know if it'll be you guys favorite. But I thought this, this was just so cool. I sent it a Minecraft skin png. If you guys have ever played Minecraft, you can essentially change your character's skin by having like a little 64 by 64 PNG image. I said, I want the outfit on this skin, just the outfit to change to blue. I want a PNG output of the same size I can put right into the game. So Manus took it upon itself to complete this task. It analyzed the skin using some sort of a command here, installed something called pillow, imported the png. Then it wrote some Python code to modify it to change it to blue, generating the brand new one, saving it as a png. Then it actually goes over and validates itself, which I like to see. I think that's something important for AI agents. I think checking over their work is very much beneficial. Then it prepares the final output and gives me the freaking skin, man. Convert it to blue. I really like that. You can actually see the two files down on the bottom hand of the screen. They're exactly the same. I could upload this right into Minecraft and it would be same skin, but just blue. It even went this far to actually make me a little diagram here. Minecraft skin modification, showing me the original and high quality on the side in green. And then the blue one, you can see the face remains unchanged as well as a lot of the other portions. I will say it doesn't modify some of the green bits over here, especially around the face and a lot of these other ones. This would be pretty easy to fix yourself and like Ms. Paint. So it did like 99% of the work here. But yeah, this, this was really cool for me. I was like, wow. You could just upload the file and it can change it for you and have it ready to go, ready to upload. Here's a replay of that session going ahead and creating that different skin. I really like that they have these replays so I can show them to you guys after the fact. But this one was pretty quick. Didn't take long. Next up, guys, we tried the martini test. This test, by the way, folks, was failed by OpenAI's deep research. It's very simple. I said, can I please get a link to buy a martini glass that looks just like this? And I sent a photo for reference. Yes, this is the martini glass. Obviously it's not a typical one. It doesn't have really a stem. It's just like a base and then a little martini thing. I don't know what you Call that the bowl. Anyways, it says I'll help you find a martini glass that looks like the one in your image. Let me search for similar options. So it searches for like martini glass, coupe style, thick stem, all these different kinds. It encountered some verification challenges, unfortunately, exploring these pages, and that's something you'll run into quite frequently. Captchkas will block these AI bots out. It's not allowed to complete them according to its programming, but I think they could probably be jailbroken by someone like Plinus to just complete captchas themselves. They have the capability to do captchas, they just aren't allowed to do them. Anyways, I suggested it to use Google Image search instead. So that's what it went ahead and did. Did some searching and then eventually provided me with some links. Now unfortunately, just like OpenAI's deep research powered by O3, it gets all the wrong martini glasses. It cannot get the same exact one. Here's another one. I mean, this is a cool martini glass, but it ain't that same one. Here's the other example again, totally different, but a nice martini glass. And here's the final one, again, nice martini glass, but not as pictured in the image. So yeah, like I said, I've tried this with plenty of AI before and they just can't seem to make it happen. By the way, here is 03 with the same exact prompt and it totally failed. We covered this in a previous video. Guys, this is the benchmark for AGI the martini glass. Find me that thing. Next up, I tried another high effort one. This was a lot more complicated. I said, could you make me a 3D animation of a hand snapping its fingers? And it goes, sure, I'll help you create a 3D animation of a hand snapping its fingers. I love how gung ho it is. I'll need to set up the necessary tools to create the animation. It might take a few minutes. So interestingly enough, it tried to install 3D animation software. It actually tried to install Blender, which is free and a good choice for doing this. It was able to successfully install Blender inside of the Linux machine it has access to, which is pretty cool. It then created a project directory, made a structure for the 3D animation, got a 3D hand model using the Python API for for Blender. It tried rigging the hand model for animation to enable finger movement. Then it was ready to create finger snap animation keyframes to define the motion of the snapping. So I was like, oh my God, is this actually gonna work? I mean, it got really far along. It moved on to adding materials and textures, improving the visual appearance, configuring the camera angles, and then finally it tried to render it. So this now is the part where it got stuck. It's a computationally intensive process that takes some time to complete. I'm like, I don't know. Can the little Linux machine that it's been provided render this animation? Anyways, as it was going on, it says the rendering process is still active and running at high CPU usage 428% which is normal. Okay, I don't think that's normal. I don't think that's possible to use over what the CPU has. I think that breaks the laws of physics. Anyways, it keeps trying, it keeps checking in on the animation rendering, but it realizes that there is a clear technical limitation within the Linux machine. If it had actually to an actual gpu, it could probably complete this task no problem, which is really crazy. I don't know if the animation would look good, but Blender requires a display connection which is not available in the current environment that it's using. So it actually went and decided to make a simpler animation using alternative methods. It found a pre made finger snap that can be customized and it actually made me a little video instead of some different various finger snapping animations which I thought was really interesting. I wonder if, if the high effort mode is like trying to find a workaround and still provide me with some sort of a result at the end of the day. Anyways, check out the video it made. Kind of scary. But yeah, it's just a green screen with some various finger snapping gifs that it got from the web. So it's something, it's something at least it provided me with a result of some form or function. I kind of like some of the GIFs. They're. They're pretty funny. I think it's funny that it found all these online and I love that it just has like some regular background music and then the subscribe please at the end. Like who, what are we subscribing to? Obviously this channel subscribe. But yeah, it definitely failed. It didn't complete the task but it got so far, it got so close and the issue was just with what it was connected to. So if you could have Manuscript hooked up to your own machine, like it could have totally done this Blender animation, at least from what I've seen from this. But yeah, this one definitely took quite a bit longer to end up giving me some sort of a result. We'll do the playback here. You can see it goes through creating the hand model and all the different crazy stuff. And again, it has like that task list that it opens itself up and then goes through down the list, completing each task and checking it off as it does. So it's so cool to see it work like this. It's very human, but also not human at the same time. It's so organized. It's. It's an AI bot, you know, it's. This is what that looks like. This is what an AI autonomous agent looks like as it goes through and tries to complete various tasks. So fun to watch it, you know, try to figure out errors and figure out workarounds to actually get something out to the end user at the end of the day. And by the way, guys, by this point, by the 3D snapping finger test, I had run out of credits. I had run out of usage, which stunk. But today I tried a little bit more. This is a 2D Python game combining Snake and Flappy Bird together. So no real specific instructions here other than it has to be in Python. Combine those two titles together in some way. So I'm really excited to see what it made. I haven't checked this out yet. You can control a Snake that moves forward autonomously, is affected by gravity, can fly upwards with key presses, grows longer when collecting food, and must navigate through pipe obstacles. It goes through. This whole thing sounds PY enhanced game. Py. All these different PY files. It seems to be pretty polished with animated clouds, apparently styled obstacles and visual effects. Complete game loop difficulty balancing that increases the challenge as you play. So to play this, apparently we have to extract the zip file it made here, ensure that Python and Pygame are installed, and then run Snake Flappy Game, PY space and up arrow key to fly and then enter to restart. All right, guys, let's download this thing it made. Alright, let's see if the game is ready to just run. Oh, okay. So we actually do need to install Pygame. Alrighty. So it wasn't too difficult. Had to install pygame ourselves here to make this actually work on our own system. But if I press this run, it opens up the Snake Flappy Bird game, which is pretty cool. So let's see. Control your Snake with the space or the up key. I already lost. Enter to restart. Oh my God. It seems like I just die immediately. Collect the food to grow longer. Oh no. Avoid pipes. There are none. This is really sad. The look game looks like it almost works. Okay, I got the game to work and now it's just like a normal flappy bird. I was able to reduce the size of the guy here, so basically made a flappy bird clone. I'm not a coder though, so it seems like, yes, you do grow, but the snake is like hitting itself every time it eats food. I think that's the. The problem we're running into with this game. So a little bug fix, but after that I think it would be able to fix itself. But yeah, it's definitely a flappy bird clone. A pretty decent one at that actually. Like in terms of the physics and the way it feels, pretty decent. I don't know where our clouds went by the way. They seem to have disappeared. But yeah, every time you eat food, snake grows too long and then collides with itself. So there's clearly a couple of bug fixes that need to happen for this game to work fully. Alright, I deleted all the code that checks if the snake has hit its own body. So maybe it'll just be fine now. Yes. Okay. The snakes getting a little bit longer. Oh, this is actually kind of challenging then. Alright guys, maybe it's like a real game now with this. Just like a couple of little tweaks. Let's get this food fall down. Yeah, it's kind of like a little bit longer. Let's see if we can get this one. Oh, it's even longer now. This is definitely difficult because you don't want to destroy the snake as it gets longer fitting through these various tubes. Oh my God. Okay. Oh. Okay. We felled. Oh man. That's actually cool. Yeah, it made a combination of snake and flappy bird. I don't know where the clouds went. We had better graphics, but they, they seem to have disappeared into the abyss, unfortunately. And some of the food is spawning in spots where it's like impossible to reach. But yeah, very simple game. Little flappy bird, snake clone kind of deal. Pretty cool. I am absolutely a fan and I am impressed that Manus was able to do this. Although I have to say I feel like Claude 3.7 is kind of the reason that it's able to do this so well. It's known for coding and it's known for being able to do it well. So yeah, this is cool. This is a really fun little demo that it made. Definitely had some, some things that needed to be tweaked. But wow, we're 46, 50. I bet you guys can't beat my score. Look at this. I am just cranking away. Oh my God. Did you see that? Dude, look at My snake. It's like floating. I am absolutely crushing this game right now, guys. Oh, there we go. Game over. Alright, 75. High score. I doubt you guys could beat that. I'm just the legend at this. In all seriousness though, I'm actually terrible at video games, so that was a miracle. Anyways, that is Manus. When this first hit the scene, a lot of people were saying like, oh, big AGI moment, get hyped. To be honest, it is the best AI agent I think I've ever used. But I think it comes down to more of just the technology aging over time. We've been experimenting with autonomous LLM based agents for a while now and as LLMs have gotten better, you know, Claude 3.703, these agents have also gotten better. But I think the framework that Manus is going for here under the hood is pretty darn good. I really would like to see them build a Windows app or a Mac app where it can take over your computer and actually use the hardware that you have to do things like actually finish that blender render, for example. I think that would be really darn cool. Now in comparison to OpenAI's operator, which I have some limited experience with, I will say an operator, when it reaches a captcha, it just makes the user take control and then complete the captcha, which is a pretty easy solution. Manus does not do that. It will try to circumnavigate the captcha in any way necessary. It'll just go to a different website or find another way to get the information. I prefer to just, you know, contact me and be like, hey, you gotta complete a captcha, just send me a notification or something like that. Be pretty easy. But yeah, this is definitely one of the most capable AI agents I've ever seen. Probably the most, honestly, if I had to put my finger down on it. Still falls into some of the pitfalls that we traditionally see though, so that needs to be overcome. Manus is cool though. I can't wait to see it when it's out of beta. I bet it's gonna be expensive though because man uses a lot of compute and that Claude API ain't cheap. Would also love to see how it does with other LLMs connected. Maybe a deep seq R1. A fully local autonomous AI agent could be cool. Completely circumnavigating that API cost. I don't know. Let me know what you guys think. Have you tried Manus? Do you have access to it and what do you think about it so far? Thanks so much for watching, folks. I'll see you in the next video, and goodbye.