You've never seen AI Agent like THIS

Hinzugefügt: 2026-05-05

[00:00:00]This right here is the space agent and it's unlike any other agent you have ever seen. It can completely change the way you work with AI agents. My name is Yan. I am the developer of Agent Zero.

[00:00:11]And today I'm here to present our newest project made by Agent Zero, the Space Agent. In this video, I will show you what Space Agent actually is and what it can do, how it operates internally, and also why is it the fastest and most efficient agent out there. Let's start with the important part. Space Agent, just like Agent Zero, is completely free and open source. You can use it for whatever you want. You can change it into whatever you want. It's up to you.

[00:00:36]Now, what makes Space Agent really different is this part right here. The agent lives in the front end. So, the agent runs in a browser, but I don't mean it has a web UI like the other agent, and the agent runs on a back end somewhere. This agent actually lives and executes inside the browser's JavaScript runtime. All of the framework, everything you see, everything the agent can do actually executes client side in the front end layer in JavaScript. This comes with some security and privacy benefits. For example, even the inference can be done in the browser using WebGPU. But the main reason to do it this way is this. The spaces spaces are a way to let the agent modify the runtime environment as it needs to.

[00:01:26]If the agent needs to show you something, it can simply draw it on screen because the agent lives and operates in the front end layer. So if you want a crypto dashboard with ticker prices, charts, and news, the agent can draw it for you on the screen in real time. or maybe a daily news overview with a list of news. You can choose different sources. You see the top news of the day and weather dashboard. Even games can be created by the agent from scratch in the browser in real time. The way spaces work is that the agent writes a renderer function for individual widgets once and this function can then be used to recreate the widget anytime on this infinite grid. Meaning even if I refresh the page, my widgets stay in place in the same shape they were before and I am not losing any changes made to the DOM. Here, for example, we have a YouTube player with a playlist of agent zero videos.

[00:02:31]>> The possibilities of spaces are endless.

[00:02:34]You are only limited by limitations of the browser runtime and your imagination. You can use it to create full applications or even systems. Now, for the agent itself, it's this little guy here. And it works like a standard chatbot. So, you can greet it. It will respond. You can switch between compact and full mode. Seeing the history, you can drag it around, you can hide it if you need more space for some of the bigger screens. So, now let's try it out. open a new space and create a analog clock there.

[00:03:15]So the agent has navigated to a new space and now it's writing code for an analog clock. We can inspect everything here. We can see the code the agent is writing. It has already executed. It's done. And we have an analog clock here.

[00:03:32]Now let's extend it a little. Let's make it a time zone check. So, we should have four clocks now. One for Tokyo, Rome, London, and I think New York was the last one.

[00:03:47]I wanted the agent to edit the original one, but I actually didn't tell it to do so. So, yeah, it's a valid result. I can remove this widget now. I don't need it anymore. And let's do one more simple edit.

[00:04:02]remove the background from it.

[00:04:11]Okay. And here we see it doesn't wrap properly. So we can tell the agent make it wrap.

[00:04:24]Much better.

[00:04:26]I'm of course aware that checking time is not that impressive, but I wanted to demonstrate how the agent operates, how it can edit existing widgets and I wanted to show how the message loop actually works in space agent because that's another big benefit of this agent. We can take a look into the message history here and see how token efficient and how very simple the communication loop actually is. So this is my first actual message to the agent. Open a new space and create an analog clock there.

[00:05:04]It only added 13 tokens into the full conversation.

[00:05:10]And all the agent had to do to get into the space took 24 tokens because the agent doesn't use any tool calling. It doesn't use structured output. It responds in plain text. So the agent starts his response in a plain text.

[00:05:24]This is the chat bubble. I can see this is the response that gets into the chat window. And then whenever it's followed by these two tokens, these are actually just two tokens. One, two. Anything after JavaScript automatically executes in the browser. So the agent used the API we have prepared to create a new space titled analog clock.

[00:05:48]This is the response the agent got.

[00:05:51]Again, we use YAML instead of JSON to save tokens.

[00:05:56]We only inform the agent that the execution was successful. And this is the result. Then the agent spends 20 tokens to navigate to the space.

[00:06:06]I'm not sure yet whether this could be done in one step or not. Maybe we can improve further.

[00:06:14]And then he writes the widget code just like this telling me rendering the analog clock analog clock widget now and following with the JavaScript of the widget render. And as you can see it didn't take much code to render an analog clock.

[00:06:34]This is the second widget being rendered. And here we can see the edits.

[00:06:40]I told the agent to remove the background and all it took was 61 tokens because the agent only responded what it's doing and then used edit function to selectively replace the background.

[00:06:56]So everything we just did is just 16,000 tokens but almost 9,000 is the system prompt. So 7,000 tokens is all it took to create and edit these widgets.

[00:07:11]Another big strength of the space agent apart from spaces is its ability to browse the web. So I can tell it go to Google and just like that the agent opens the Google window. I'm sorry that this is not English, but it's it's probably using my IP address to determine the language. So, I can just tell the agent accept cookies and it should have no trouble identifying the button and pressing it. Now I can tell it search for agent zero and open the official web.

[00:08:10]And just like that, this is our site.

[00:08:15]Accept cookies again.

[00:08:19]And here we are. Now let's give it something that will require multiple steps.

[00:08:25]Now navigate to the GitHub repo and find the oldest release there.

[00:08:48]The oldest release is version 0.2 which is correct. And the best part about this is that I can interact with the browser in real time just like the agent does. So if there is a capture, if there is a login screen or whatever, I can do it and tell the agent continue from here. Also, let me show you how efficient the web browsing in space agent is. If I go to the history again, now we only have 22,000 tokens.

[00:09:15]If I'm correct, previously it was 16,000. So all this web navigation added just 6,000 more tokens. And that is because we don't leave we don't leave any leftover data in the history.

[00:09:31]When a website is opened and we parse it, we show it to the agent like this.

[00:09:37]We basically transcribe the whole DOM tree into a list of images, links, text, etc. So the agent can see and can reference anything on the website, can click anything, it can scroll, navigate, etc. But this message lives in the transient space of the prompt. Meaning it's always appended to the very end of the prompt after the last caching break point. So it doesn't break caching and it saves a lot of tokens. Another big advantage of space agent is the customizability and extensibility. The agent lives in a virtual file system that is accessible by the front end. This is my user folder with my spaces being stored here in YAML files.

[00:10:27]And the whole framework is designed to be extensible by default. Everything in the framework including the core is built as a module and modules can be added or removed at any time and they can be developed by the agent itself.

[00:10:45]The agent is skill-based meaning everything the agent can do lives in the file system in form of skills.md files.

[00:10:54]The browser, the spaces, the development documentation, everything is in skills.mmd files. So it's easily extensible and the agent can develop itself further.

[00:11:07]Right now we have a few basic functions here like the file browser. We have a testing area for local models if you want to run the AI inference locally in the browser.

[00:11:20]It works but requires a beefy GPU.

[00:11:24]I don't expect the speed to be great here. And modules like this can be developed by the agent itself at runtime. You don't need to do any changes to a server, restart it or whatever. It can be developed completely on the front end. And the best part is that all the changes are scoped in this framework. Meaning you can have a multi-user system and users can develop inside their home directories completely new functionality without affecting others. Or you can do the same for user groups. If you want multiple users to share the same functionality like an accounting department in a company, you can develop something from them, put it to a group folder and it will be shared across these users. They can have read access or write access. It's up to you how you configure it. This separation also allowed us to implement time travel. Every user folder or every group folder has a G repository automatically created in it. And I can travel back in time and undo any of my changes. Like for example, here somewhere here I created the new space with the analog clock. I can travel back in time before that.

[00:12:38]And my analog clock is here containing only the original widget. If I go to the time travel and travel back even more probably somewhere here, it should be gone entirely.

[00:13:00]And I can also travel forward back to present.

[00:13:06]and my workspace is as it was.

[00:13:10]So this way you don't need to worry about breaking anything while developing the space agent itself. You can always fix it simply by traveling back in time or reverting individual changes in your file system or if something really breaks and the page doesn't even render. There's also this admin mode which splits the screen in two.

[00:13:35]This one is static. It is built into the firmware of space agent. And this will allow you to browse the files, talk to the agent, and time travel even when the page itself completely breaks for you.

[00:13:49]One of the biggest benefits that come naturally from the fact that the agent runs in a browser is that you can run the agent in a browser. So you don't need to install anything. If you want to try space agent now, you can go to our repo and click the try live now button or you can go directly to space-agent.ai and here in your browser you can create a guest account with a single click.

[00:14:16]login with another click and you are right in. No installation, no configuration except for setting the API key for your LLM or using the local inference if you have the M machine that can do it. But that's it.

[00:14:32]You're right in a working space agent instance. This one is a demo site. These accounts are not permanent. They will be deleted in a few days if you stop using them. So if you want to use space agent persistently and you want to make it your own, you can go back to our GitHub repo and you can download it as a native application for Mac OS, Linux and Windows or if you want to run it in a browser, you can host the back end yourself.

[00:15:05]There's a guide for it. There is a very thin backend layer in NodeJS that only serves the files for the front end and manages user permissions and user accounts and low-level stuff like that.

[00:15:17]The actual framework and all of its functionality runs on client side in the browser. Obviously, this is just the beginning. We have very little idea what's actually possible with space agent. I believe people will create wild things. If you create something interesting and you want to share, there is a button for it. It will allow you to download and upload a zip. But you can also share it through our website by creating a share link. And you can copy this and share it with your friends. And they can open it in a browser just like that in a sandboxed environment. So it's not in their application, it's not in their environment. So they don't need to worry about malicious code stealing their secrets and they can get into the space just like that to test it. As always, thank you very much for your time and support. Please consider giving us a GitHub star and joining our community on Discord, school, subscribing on YouTube. Everything helps. Thank you very much. See you next time.