LiveSpace: Making the Frontend Come Alive
Jase Broderick
SubscribeSign in
LiveSpace: Making the Frontend Come Alive
Jase Broderick<br>Jun 18, 2026
Share
Previously, we launched AgentSpace, a new distributed coordination substrate that allows for a shared memory association across various processes. Today we are looking at one of the applications of AgentSpace, which we call LiveSpace. This application allows for front-ends to launch into a new paradigm of state computation. We believe this is needed, as current generative UI tools are hamstrung by the inability for the frontend to react quickly and lively to new web components, not coded beforehand. In this post, we explore how we can use AgentSpace on the frontend to create new paradigms and UIs using LiveSpace as a substrate for front end state computations.<br>Frontend History: Stuck in React
The expansion of the frontend has changed over the three decades since the internet was launched, moving away from simple javascript webpages, to huge sprawling engineering projects today. The current front-end paradigm is best encapsulated by a front-end framework called React. React is built on the principle that the UI is essentially a state, a certain data model that needs to be computed. To compute that state, React allows for the use of hooks, to allow for actions from the webpage to come in and be used in different services. While there are other web frameworks, none try to steer far away or challenge this central idea of the state. This fundamental assumption about how to set up the state computation is the core of front end development since the 2010s.<br>While today the web is very much a jungle of various React projects, it wasn’t always this way. Previous web iterations included things like Flash animation, a web paradigm that allowed for more expressive and richer web content within the browser. While Flash eventually died out, the need for expressive and reactive web frameworks never really went away. With new generative AI technologies, new approaches to user experience have been executed/implemented trying to generate these web UIs on the fly. Google recently tried to generate full web pages from large language models, but these generations are limited to prototypes today. (https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/ )<br>Currently, the biggest consumer applications of GenAI include chat applications (ChatGPT, etc.); all of these applications have simple generative UIs needs (text) during the chat session. Recently, these chats have started to include things like documents, HTML files, and animations. While these advances are great, we believe they are a far step away from what could truly be done today.<br>Generative UI Infrastructure: Using AgentSpace to Store State and Publish it to the Frontend
In order to reach what we believe is the maximum potential of this new technology, we are introducing LiveSpace as a new method of calculating state. Instead of having the state coupled by the DOM, the state can merely be a tuple that is stored within AgentSpace. Using LiveSpace, the library will handle storing and creating the state in AgentSpace. Then, a backend Agent Application will put these states into a tuple to be computed later by the frontend. Since this state can be stored and taken from AgentSpace, we can do extremely fast live updates.<br>The following is a stress test of the system: 100 numbers and a live list all updating at once. What we want to show is where the state lives. Instead of each component holding its own state, every value sits in the space as a tuple, and each component just renders the latest version it gets handed. The source of truth is one shared log, and many producers, including agents, can write to it at the same time. The stress test is what that looks like when a lot of those values change together.
While the UI Infrastructure can be used to have a large amount of updates go through the browser, we can also use it to quickly generate or bind components on the fly for live generations of components. Take for example e-commerce, today static e-commerce websites will need to use a typical feed design to showcase their SKUs. With LiveSpace, we can create more interactive experiences to have the webpage be more interactable and snappy to the user’s intent. Here, we have a demo shopping experience, instead of a typical feed, we can contrast SKUs on a pane, then after collecting user feedback, the comparison UIs can be generated on the fly:
In our final example, we imagine how front end components as a whole can be generated on the web. In this example we show how math visualization can be generated on the fly, users can prompt what math concept they would want visualized, and then it will be placed onto the web page, allowing for new custom generations of software to come in for educating the user.
With these new additions to the AgentSpace framework, we believe...