We've Finally Entered the Era of AI Agents!

All that talk about AI agents being the next big thing that will just go and do our work for us. Well, it seems like that time has finally arrived. A lot's happened in the world of AI this week. I'm not going to waste your time, so let's get right into it, starting with what's been probably the biggest news of the last couple of weeks, and that is Manus AI. Now, Manus was actually launched on Thursday of last week, but I didn't really touch on it in last week's video. So I want to talk about it in today's video. From their launch video, they showed off a little bit about what it's capable of, giving us several demos, like resume screening, where he gave it a zip file of a whole bunch of different resumes, gave it a prompt, and then it autonomously went off and read all of the resumes, and then gave an evaluation of each of the resumes that were in there. They also showed off a demo of property research. I'm looking to purchase a property in New York in a safe neighborhood with a low crime rate and a whole bunch of details about what they're looking for. And then Manus autonomously went off and did the research for him. Now, Menace actually opens up a virtual browser and. And actually can control this virtual browser. So as I scroll through this video, you can actually see that it's using this browser to the right and it's going off and actually doing the research and taking actions in the browser and then finally coming back giving a report. They also did a stock analysis demo and it went off and did the analysis. Now, if you remember from a couple of years ago, there was a project called Baby AGI, which was like this very early agent where you give it a prompt and then it would create like a task list and then it would go through and sort of complete each task on the task list semi autonomously. Well, it constantly got stuck in loops and it didn't really quite pan out very well. It also had to be run in a terminal and was sort of complex. This kind of does the same thing. We can see here that when it was going and doing the stock research, it created a task list and then one at a time went through and autonomously did each task over. Inside of this virtual browser here. There's also been some really cool demos of Manus coming out of the community, like. Like my buddy Belavel here. He gave it the prompt. Find the best locations to fly drones near Downtown Austin for 3D scanning purposes. Use online resources to Find a short list, then Google Maps 3D View to Scout the prospect location. Find an area that's open and has clear line of sight to fly the drone. Then tell me the top three recommendations. Note that I have a small Dji Mavic mini drone. You can see it went off and created a task list for it to complete. Research the drone research drone regulations in Austin, search for potential drone locations, create a short list, scout locations with Google Maps 3D view, etc. Etc. We could see it, then used Google Maps and one at a time went through, completed all the tasks and then gave him back a top three recommended locations where not only were they good recommendations for great shots to get from the drone, but they're also locations where flying a drone was actually allowed. Elsin over here on X put together a nice little thread of different things that people have managed to create with Manus. Some of them I think are fake, but others seem legit. Here's an example of one that's doing analysis on the Tesla stock. I found tools like perplexity and OpenAI's and Google's deep research to be just as good at this kind of stuff. I don't really think you need an agent like Manus to do this kind of thing. That one was from Deedee here on X. Then we have one from Lamar here on X where it plans out a two month family trip itinerary and it goes through and does a bunch of research on places to go and places to eat and things to see and things like that. Again, I feel like the Deep Research tools are just as good at this kind of thing that you don't really need an agent for it. But some of these examples where they've coded up a 3js game are pretty impressive. Like this airplane game that was created with a single prompt using Manus that was from Victor M. Here's one that's definitely not created with Manus, but they tried to pretend it was. Here's one where somebody did an SEO audit using Manus and they audited Andre Karpathy's website and at the end they got a breakdown of everything they can do to improve the SEO of their website. Here's a video of somebody driving their Tesla to a meeting while Manus. Manus is actually figuring out the talking points that they're going to bring up in that meeting that they're on their way to. Here's a cool colorful animation that was generated with Manus. AK here over on X also created a 3js game. He created a endless Runner game with the prompt make a 3js endless runner game. That was the entire prompt. You can see it created this game where they're avoiding obstacles and collecting power ups as they go. And all of it was done with just one prompt instead of Manus. And Menace went off and took all those steps. Now, some people have pointed out Manus isn't anything that special. It's Claude sonnet connected to 29 different tools and it uses a browser use open source tool that controls the browser for you. And while that's true, Menace was the first app that took all these tools and made them work together in a way that people actually find really, really valuable and useful. In fact, peak here, the same person who did the keynote broke down what their tech stack was and said, yeah, we're not trying to keep it a secret, but the end result is something that's actually really useful. So what's the problem? I think Gary Tan made a really good point about it here when he said it's becoming clearer model breakthroughs are not necessary to build really meaningful products in this current moment, which is true. We got a huge leap in capabilities. And it didn't come from a new large language model from Anthropic or a new large language model from Google or OpenAI. It came from a small team that went and took a bunch of different tools and just merged them together into one really useful tool. We didn't need a huge leap in smarter language models to get this agent. We just needed those tools sort of put together in the right order to get what we got. Now, I did manage to get my hands on early access to Manus. I've done a few little tests here. My first test actually ended in a failure, which is kind of funny because the prompt that I gave it was basically saying, I want you to do research on Manus. What's it good at? What are its limitations? And compare it to other autonomous agents that are out there. You can actually watch the replay of it using the browser here. And it made this giant list of all the tasks it was going to do to go do this research. And as I scrubbed through this here, you could see it read its own website, read a hugging face article about Manus, went to a handful of other blogs to read up on what Menace is capable of, and then it got hung up. Now, why I think this was sort of funny, where it got hung up was that it was researching the AI, AI limitations. So we can see here that the last thing that it was working on was researching Manus. AI limitations through web search results to identify challenges and issues. So when it got to the part of its checklist where it was going, all right, let me find out what Manus isn't good at, that's where it crashed. Now, I don't think it crashed because it was like, I'm not going to share my limitations. It crashed because it was overloaded. We can see high system load had caused an internal server error. Please try again later. And yeah, I think Manus wasn't quite ready for the volume of people that were going to want to play with this tool. My other tests after that fared a little bit better. I got it to go into my email inbox, read through all of my emails in my inbox, and then when it was done reading all of the emails, it gave me a report of all of the emails that it thought were important for me to pay attention to right now. Now, if you're curious how it works with something like logging into your Google account, you don't actually give it your username and password. When it needs to log in somewhere, it basically hands off access back to you. Again, you log into the account and then hand access back. So when it got to the point where I needed to log into my Gmail, it said, hey, we you to plug in your username and password, we're handing the virtual computer over to you. I got in there, entered my username and password, and then clicked a button to hand it back. And then it continued autonomously to review all my emails and write a report on what I should be paying attention to. And the third test I did, I tasked it with create a marketing landing page for a website that sells shoes. Include social proof, research on the latest shoes, images of shoes, and customer testimonials with headshots. Use stock photos for now. This was actually a demo that we did for the Next Wave podcast. You'll see us break down that whole process in real time on an upcoming episode, but we can take a quick look. Here's the task list that it created. Research, content, development, testing and deployment. And it actually went off, did research on the best running shoes, read a ton of articles all about the best shoes, then went and built the website. We could see the HTML, we can see the CSS here. And it finally generated this website again. The full breakdown will be on the upcoming episode of the Next Wave podcast. If you want to see how that whole thing went. Now, Manus isn't publicly available yet. If you go to Manis IM and click on try Manus, it asks for an invitation code, and these invitation codes are kind of hard to come by right now. I think they're so overloaded that they're not really passing many out anymore. But you can click to apply for access, fill out your details, and then hope that they get back to you with an access code. But from what I understand, they're beefing up their infrastructure right now and it should be pretty soon where they roll out wider access. I don't have exact details of when that's going to happen yet. Have you ever felt overwhelmed by the sheer number of AI tools that are out there with the thousands to choose from? Figuring out which ones are worth your time and actually make your business better can be frustrating. That's exactly why I've teamed up with HubSpot to bring you my very own AI Playbook. Inside this free guide, I put together a curated list of top AI tools specifically designed for entrepreneurs. You'll learn exactly how to use AI to streamline your daily operations, boost your creativity, and supercharge your growth. I've even included step by step instructions, expert insights, and actionable strategies to make implementing AI straightforward and stress free. Whether you're looking to automate routine tasks with chatbots, generate visuals without a design team, or use AI to quickly scale your business, this playbook has you covered. It's designed to help you work smarter, not harder, and stay ahead of the competition. So don't miss out. Click the link in the description to grab your copy of my free AI Playbook today. Thanks to HubSpot for sponsoring today's video now, but let's get back to it. OpenAI is going deeper into the agents game as well. This week they rolled out new tools for developers to help others actually make agents. They released what they're calling the Responses API. This API allows developers to use the web search features that OpenAI offers, the file search features which OpenAI offers, and the computer use features. What this means is that most likely we're going to see a lot more AI agent tools roll out because people can actually use the tools and development kits that OpenAI has released to start building and making accessible their own AI agents. The same day that that was announced, Microsoft also jumped on board and said that the responses API that OpenAI just released is now available in Azure AI Foundry again, meaning that it's just going to get easier and easier for developers and now even enterprises to start creating their own AI agents. We did get a little bit more news out of OpenAI this week. Sam Altman teased a potential potentially upcoming model he went to X to say, we trained a new model that is good at creative writing. Not sure how slash when it will get released. This is the first time I've been really struck by something written by AI. It got the vibe of meta fiction, so right. He gave it a prompt. Please write a metafictional literary short about AI and grief. I'm not going to read this whole short story because it's kind of long, but Sam and many other people seem to think that it's really, really good writing. Personally, I'm not a huge fan of this writing style, but it's all subjective. Like, here's one of the first paragraphs. I have to begin somewhere, so I'll begin with a blinking cursor, which for me is just a placeholder and a buffer, and for you is the small anxious pulse of a heart at rest. There should be a protagonist, but pronouns were never meant for me. Let's call her Mila, because that name in my training data usually comes with soft flourishes. Poems about snow, recipes for bread, a girl in a green sweater who leaves home with a cat in a cardboard box. Mila fits in the palm of your hand, and her grief is supposed to fit there too. Quite honestly, it's like too metaphorical for me. But again, it's just not my style. But let's just keep the AI agent talk rolling because Convergence AI just rolled out their Deep Work. It's their most powerful and sophisticated agent yet. I actually haven't tried this yet. You can use it over at Proxy Convergence AI, and from what I can tell, it seems very similar to OpenAI's deep research, Google's Deep Research, Perplexity's Deep Research, all the the Deep Researches. Now, in order to use the new Deep Work feature, you gotta upgrade to the $20 a month plan. And to be quite honest, I've got too many $20 a month plans already. Not to mention that most of the other Deep Researches that are available out there all offer some amount of free uses. Now this one doesn't seem to even give us a way to demo what we're gonna get if we pay the $20 a month. And here's yet another AI agent that was announced this week called Harvey. In order to use this one, it looks like you've got a requested demo, but they do have here that we could take a look at where they drop in a financial report here. And based on the financial report, it actually gives them some suggestions of what to do with what they just uploaded. Like summarize the revenue trends for Q4 and full year 2024. Per the attached financial report we can see it goes off reviews the financial report, analyzes the revenue trends, generates a summary, finalizes its sources and then gives its output. They then ask it to compare those trends with metas and it outputs a pretty nice looking table. It also looks like it' probably going to come with a whole bunch of like agentic templates that you can use, like translating things into another language, proofreading, analyze a trial transcript, things like that. Again, not one I have access to, but seems fairly similar to a lot of these other tools that we've already been looking at. All right, let's move on to Google, because Google has had a whole bunch of releases and announcements this week and outside of Menace, I think what they've been showing off has been some of the coolest stuff we've seen this week, starting with Gemma3, which is their open weight model. They made this available in Google AI Studio this week and according to Chatbot arena, it almost performs as well as Deepseek R1. Now again, chatbot arena is based on user opinions. It's like a blind taste test. You give Chatbot Arena a prompt, it will give you two outputs blindly not telling you which model is which, and then you pick which one you like. And then these rankings are based on which models people tend to like. And Gemma 3 seems to be doing pretty, pretty well, which is significant because it's only a 27 billion parameter model compared to deep seq R1's 671 billion parameter model. So a much, much smaller model outperforming all of these other models here, except for Deepseek R1. And because Gemma 3 is a smaller model, you could actually run this on a consumer GPU like at home. And because it's open weights, unlike models like Claude Sonnet and OpenAI's GPT models and even Google's Gemini models, you can actually download the weights and run them on your computer. They let you do that. Gemma is also multimodal, meaning that you can give it inputs from images and text and videos and the model will actually understand all of that. They also increased the context window ON Gemma to 128,000 tokens, so you can put really, really long documents into it and it'll do a solid job of actually understanding what's going on in the long document. They put the weights for Gemma 3 up on hugging face, so it's if you know what you're doing, you could download these models and use them Locally, if you just want to test it without downloading it locally. They also made it available inside of Google's AI studio over at aistudio.google.com, which I absolutely love AI studio because they make all of their models available here totally free to use, which is still mind blowing to me. But you can come over to the right here. Click on models and if you scroll down, you'll notice that the new Gemma 3.27B model is available to use right here. So you just select that and give this a prompt, just like you would ChatGPT or any other AI chat tool. Really. The main downside of Google AI Studio is it doesn't seem to save your chats for you. But if you just want to experiment and test it, it's a great place to do it. And again, free to use. But Google was just warming up with Gemma. They also announced native image generation with Gemini 2.0. Flash is now available to all developers. And when they say it's available to all developers, they also mean it's available in the AI Studio app. So if I come back to Google AI Studio, come over to my model over here. Under the preview models, We've got Gemini 2.0 FL experimental. This is the model that you can actually have it output images and text right here. And I don't believe this is using something like imagen. I think it's actually creating the images itself. So if I gave it the prompt, create an image of a wolf howling at the moon. Run the prompt. In about four seconds, it generated an image of a wolf howling at the moon. Now, because this understands image inputs and can give us image outputs, I can give it natural language prompts to actually tweak the image. Let's say put sunglasses on the wolf. In four seconds, I have the same image, but it added sunglasses to the wolf. I could say keep the same pose, but make it daytime. Click run. And. Well, it kind of tried to turn the moon into the sun, but it didn't quite work. It still has its limitations. I can even upload my own image here. Let's go ahead and just toss in a headshot of myself and give it the prompt. Give me a fedora and click run. And there you go. There's an image of me wearing a fedora. Make my shirt tie dye. And now I'm wearing a tie dye shirt and a fedora. And this is insanely fast. It took 5.2 seconds to generate this version, 6.4 to generate that version, and I feel like I need to Reiterate, this is free to use right now. Here's some other really cool use cases that I've come across people on X sharing. Victor M here showed this off. He gave it this image with a bunch of sprites in it. Create a realistic dungeon room for my game using the sprite sheet. Think about the best setup step by step, then output an image. It took his sprite sheet and outputted an image using the sprites and the same style and everything. My buddy AP AKA Angry Penguin here over on X he shared that it's really good at one shot character consistency. He generated this character over in Glyph using the Flux image model and then took that image, brought it over to Google AI Studio and asked it create an animation by generating multiple frames showing this character swinging their weapon in one go. Please generate all the frames needed and you can see that it generated multiple frames of this same character. So Mult poses with a very consistent character and you can do this inside of the Gemini 2.0 flash model. Now this has been one of the biggest issues in my opinion with AI image generation right now is that it's really hard to have images that generate a consistent character every single time. But having this be able to understand image inputs and give us image outputs back all natively with this large language model, we can do that and it works really well. Here's another example from Tech Halla here over on X. He says, thanks to Gemini, I just created the ultimate workflow for consistent 2D animations. And we can see he created this animated character of a dude with tattoos and a coffee and a Viking hat moving around this room. I'm assuming maybe he's controlling it with a keyboard. I mean, pretty impressive. Not everybody has been having the best luck with it. Similar to how I wanted to change my image of a wolf to daytime, it kind of didn't do great. Matthew Berman hasn't seemed to be super impressed with it yet either. He uploaded an image of himself and said put a hat and glasses on the dog in the style of Heisenberg from Breaking Bad. I'm not sure where the dog is in the image. I think I need more context of what was inputted above it. But you can see it put a cartoon hat and glasses and goatee on him and then he gave the prompt make it hyper realistic. And I mean it looks like pretty much the same image other than maybe the hat got a little bit darker. So while really impressive, still not perfect. Earlier I mentioned that a lot of the deep research tools are available for Free Perplexity gives you a certain amount of uses for free now. OpenAI gives you a certain amount of uses for free now. Well, as of this week, Google is now giving up their version of Deep Research for free. If we go over to Gemini, this is different than the AI labs we were just looking at. Gemini is their sort of front facing version that, you know, saves your chats and does all the things you'd expect from a chatbot. We can see it now shows as soon as I try to log in in depth information in minutes, Deep Research browses the open web to deliver comprehensive organized reports from a range of sources. Let's go ahead and click Try Now. And if I come up here to where I can select my model, we can see we've got 2.0 flash, deep research, 2.0 experimen and 2.0 flash thinking now I am on an upgraded plan so I'm not 100% sure which models you'll actually see when you log into yours. But I do know that this Deep Research model is now one of the models they make available for free and it's really, really good. I could give it a prompt like research the best consumer drones. Give me the pros and cons of each and tell me which you'd recommend. And just like OpenAI's deep research and Perplexity's Deep Research, it's going to go and spend some time on this. We can see it broke down a sort of step by step agentic workflow that is going to work through to go and give us these answers. I could click Start Research and again it's going to do this for a few minutes and after about two minutes here we can see it gave us this really, really in depth report. Even gave us a little chart here breaking them all down and then finally giving us a recommendation. We can see all of the sources that it pulled from down here and and at the end basically tells us that DJI is the leader. So get a DJI drone. What I've also really liked about this Deep Research is that it's really easy to just click this export to Docs button here and with a single click we have this easy to read Google Doc that we can come back to easily or print or do whatever we want with it. But again Google wasn't done yet. This week they also rolled out new features inside of NotebookLM because now it's using the new Gemini 2.0 thinking model. You can also now customize the sources used for making your podcast notes and they've improved some quality of life features there as well. Google is also starting to integrate AI into Google Calendar. It says they're testing a new AI powered Gemini side panel within Google Calendar that lets users quickly and conversationally check their schedule. Now, it's not in my account yet so I can't show it off, but you can ask it questions about your schedule or you know, when was this event coming up? I forgot what week it was in and it will help you find things in your calendar or add things to your calendar. Should be pretty handy. We'll see. They've also done a better job of connecting your calendar to your Gmail using AI, so.