Transcription
So, this is Manis, a generalized AI agent system that allows you to build, or research, or do pretty much anything. And it's using the most cutting-edge tools out there. It's like combining Operator with Deep Research, giving it its own Linux environment, having multiple agents work together. And it works really well. Let's just watch a demo of how it works.
So, this is deeply analyzing the Tesla stock, and this is a replay. So I'd like a thorough analysis of Tesla stock, including—and we give it a bunch of instructions. The first thing you'll see it do is come up with a to-do list. And so this is what you're seeing, and it's operating at the same time, by the way. So, create a project directory, create a to-do file, outline all the requirements. And it's really cool because you can actually see it working along the way, and it can use the browser, and you can take over at any time and check out the Linux environment. It's super impressive.
But the cool thing is, most of these tools are open source. It's probably using CLA 3.7, Sonnet under the hood. It's using BrowserUse for the actual agent to be able to control the browser; it's using Linux. And so technically, anybody can really create this. Not to say that it's not really difficult to do it right, but they're not using a fine-tuned model; there's not really anything under the hood that you couldn't use yourself. But again, it's very, very impressive and very difficult to pull off. Well, so here we can see, I'll help you create a comprehensive analysis of the Tesla stock. Here it's connecting data sources. So, of course, it can do web search. Here you can see it's navigating through its own Linux environment. So, make directory P Tesla analysis and CD into that directory and create the to-do file. And so this is definitely the most comprehensive no-code agent software that I've ever seen. Here you can see it navigating through the directory; it's installing Beautiful Soup, and here it's starting to output some information. Now, here if we go forward, we can also see here's some code, here's the terminal; it switches back to its to-do list and checks it off as it's going. I mean, it is just super impressive.
I'd say the most impressive part is its ability to make errors and then fix those errors. And so when you kick off a task in Manis, you should be prepared for it to be going for a while. I asked it to write two games so far, and it took hours to do. And so look at all of this output; this is all the research that it's doing. So you could see it's actually creating different charts and graphs, and at the very end, we actually have a live dashboard. So if we click into it, here it is. This was created by Manis all in one go. So here you could see the investment recommendations, the different financial performance and valuation analysis, market analysis, technical analysis. I mean, this looks absolutely gorgeous and was completely built by Manis.
So they have a bunch of different examples. Here's one: Trip to Japan in April. So we'll skip to the results; it's just replaying one that's already done. And so here's the 7-day itinerary: Day one, arrival in Tokyo; Day two, exploration; Day three, hidden gems. Really, really impressive, and it did a bunch of research on our behalf. And so it does have a benchmark; it has the Gaia Benchmark, which it's only comparing against Deep Research, which I found interesting. And so we have level one, level two, and level three, and in all three, Manis passed at one, scored well above Deep Research; level two, about even; and then level three, well above Deep Research. Now, again, I find it very interesting that it's only comparing against Deep Research because it is so much more than just Deep Research, since it has its own environment, its ability to code, its ability to execute code; it can go browse the web, similar to Deep Research, and just so much more.
Now, Jen on X says, "So I simply asked Manis to give me the files, and it gave me their sandbox runtime code. So it is claudonet with 29 different tools," it says, "without multi-agent, but I'm pretty sure it's using multi-agent." I think there's a manager agent, and I think that manager agent dispatches things to specialized agents to accomplish it. It is using BrowserUse, which is an open-source framework to allow agents to control a browser. And the CEO of Manis has actually jumped in and confirmed, "It's not that complicated. The sandbox is directly accessible to each user; each session has its own sandbox, completely isolated from other sessions; user can enter the sandbox directly through Manis' interface." And again, Manis is impressive because it is the most holistic use of AI that I've seen. It basically takes the best of everything we've seen out there and puts it all together in a really nice package. The code in the sandbox is only used to receive commands from from the agent, so it's lightly used. The tools' design isn't a secret, and it says right here, "Multi-agent implementation is one of Manis' key features." When messaging with Manis, you only communicate with the executor agent, which itself doesn't know the details of knowledge planner or other agents. I mean, this is just so, so cool. And we did use BrowserSource's open-source code, which is great. It's open source; that's the point. In fact, we use many different open-source technologies, which is why I specifically mentioned in the launch video that Manis wouldn't exist without the open-source community. We'll have a series of acknowledgements and collaborations coming up. I truly hope Manis itself goes open source, but I'm not holding my breath. But towards the end, he does say, "We'll be open-sourcing quite a few good things in the near future."
Now, although it's being reported that they're using Claude Sonnet, he does say right here, "We have a lot of fine-tune models in our company; some of Manis' planning techniques originated from this model I've open-sourced last October." So there it is, the Steiner 32B preview. Thanks to the the sponsor of this segment, Langchain; they have been awesome partners to us. So excited to tell you about them again today. Langchain is a leading AI software development consulting company that builds AI products to propel your business forward. Those products include an open-source and open Telemetry-based observability and evaluations platform that helps you evaluate and improve your LLM usage in your application. It's trusted by thousands of developers from early-stage companies all the way up to Fortune 500 companies. Langchain helps developers collect and analyze traces, collect data sets, and run evaluations, resulting in highly reliable and secure AI systems. Again, Langchain's open source and open Telemetry and plugs in easily with OpenAI, Mraw, DeepSeek, Gemini, Weave, Pinecone, and more. Langchain offers end-to-end observability, tracing everything from LLMs to vector databases and framework-level calls like Crew AI, LlamaIndex, DSPI, and Langchain, with native support for Crew AI, which you know I love. Langchain provides a custom-built dashboard to track Crew AI sessions, agents, tasks, tools, and memory; track everything your agents are doing. So go from shiny demos to reliable AI products easily with Langchain. Check out Langchain; it is open source, and if you want to use their hosted version, you can get 20% off right now if you use the link in my description. And if you want to learn more, join one of their coming webinars where they go over everything. So check them out; they've been a great partner. Go star their GitHub, and thanks again to Langchain. Now back to the video.
All right, so let me show you a few examples of what people have built with Manis. So this is AK from Twitter prompt: Make a 3JS endless runner game. And there we go. I mean, this is super impressive; this is one prompt, one go. Here are some more examples. Here is: Clone the Apple website. And this sure looks good to me. And of course, here is the 3D browser-based flight simulator game, and this looks darn good as well. Here's an AI game; kind of looks like Red Dead Redemption, but you can't walk around yet, but still very impressive. Here's an incredible animation; really cool, really dizzying. And here's another open-source project that we know it uses now, E2B, which is an open-source coding runtime for AI agents. So a lot of really cool open-source technologies all being put together in a single, very polished product.
I was able to get a Manis invite code from Hattis from Groo. Thank you so much to her for giving that to me. And one thing that surprised me is that Manis is actually the same company that made Monica.IM, who has sponsored this channel, although this video is not sponsored, and I got access to Manis on my own. So I created two different games just to try it out. The first is one that I've seen everywhere online, which is: Create a 3D browser-based multiplayer flight simulator. And I wanted to see how far we can get in one go, so that's what I did. And you can already see, so making the directory, making the to-do, and if I click into it, you can see the to-do list: List, so research 3D web technologies, research multiplayer implementation, evaluate framework options. And it took a long time, but of course it should; it's a big project. And so you can see all of the output here, and at the very end, it actually tried to run the game within its own environment, but I couldn't get it working. Uh, so I actually just said, "Download me the files," and surprisingly that worked; it zipped up the files and let me download it quite easily.
The other game that I created is a game called Choo Choo Rocket. Maybe you've heard of it; it was a really fun game way back on the Dreamcast console, and it was kind of a unique game where there's these little mice running around a board, and you basically try to get them to—if I remember correctly—fall into a hole, and it's a multiplayer game, kind of high intensity, high energy; it was a lot of fun. So I asked it to create that as well, and it actually did its research, which you can see here. So it found a PDF for it, and it scrolled through the PDF, learned everything about the game, and this is the game. It's kind of hard to read on this PDF, but it's really cool; it's basically the exact instructions on how this game works. And so then it started building. "I've successfully implemented the core game mechanics for our Choo Choo Rocket web implementation," and all the way at the bottom, it finally did it, and it told me how to get it running and allowed me to download the file as well. Now, unfortunately, both of the games that I created out of the box didn't quite work; they were kind of close to starting, but they didn't work. So here's an example; you can actually see the 3D flight simulator kind of loaded up, but definitely not working. And here's Choo Choo Rocket: Enter your name, I'll join a room, hit ready, and it says, "All players are ready, game starting," but does not start. So I could probably work a little bit more with Manis to get these working, but I just wanted to show it to you and show you that sometimes it doesn't get it going right away.
And already, because a lot of the underlying technology of Manis is built with open source and nothing really specialized, we we have multiple open-source versions of Manis already. The first is called Owl by Camel AI, and here it is; it already, by the way, has 8 1/2 thousand stars; it achieves a 58.8 on the Gaia Benchmark, and that was within just a few hours of being created. And so just like Manis, it has real-time information retrieval, multimodal processing, browser automation, document parsing, code execution, and a ton of built-in tools. And then we have another version called Anis—yes, that is how you pronounce it—Anis. Same thing: Execute complex tasks, collaborate in multi-agent environments, execute with web services. And this is so awesome. Every time a new project comes out where it's closed source, although we'll see if Manis ends up being open source, the open-source community dives in and starts building it. And the funny thing is it was actually built with Manis, so Manis was used to build one of these open-source projects. So Manis is very impressive. I've seen a bunch of really good working demos, obviously on the website; you can see it yourself. So yeah, I'm going to keep playing with it. Let me know what you think. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.