Most "AI-powered" apps today are a regular app with a chatbot bolted on the side, pointing at the real buttons from a distance. The chatbot can describe what the app does but usually can't do it, and when it can, it's taking a separate path with its own bugs. Builder.io thinks that whole setup is backwards.

The Bot and the Buttons Now Share a Brain

On October 14, Builder.io opened the beta of Agent-Native, an open-source TypeScript and React framework for building apps where people and AI agents use the same actions, data, and application state. You define an action once, and the framework exposes it to the visual UI and to an AI agent. It also exposes it over HTTP, the command line, MCP, and A2A.

According to Builder.io's launch materials, the framework includes auth, permissions, and real-time collaboration, and it works with any AI model. The visual interface runs alongside agent chat, so users can switch between clicking and asking without the two falling out of sync. Builder.io also runs hosted versions at agent-native.com, but the framework is meant to support teams that deploy and run their own apps.

Why "One Action Layer" Is a Big Deal

When an agent bypasses your UI and calls the database through some side channel, every business rule you put in the front end, from validation to permission checks to the "are you sure?" prompt, gets skipped. A shared action layer means the agent goes through the same checks a human does. That's better architecture, and it also makes security much easier, because one permission model is far easier to audit than two that are supposed to match.

Here's the catch. Coverage notes that adoption and production reliability are still unproven, and an open-source release can also be a way to steer developers toward a vendor's hosted service. You don't have to adopt this framework to use the idea, though. Whether you're on Next.js, Drupal, or something your team built a decade ago, exposing your business logic as well-defined, permission-checked actions (ideally over MCP) is the best way to prepare an existing platform for agents. Teams that do that work now can plug agents in later without having to rebuild.

The chatbot era is giving way to the agent era, and in the agent era the AI uses the same controls your users do instead of guessing at them.

If you want AI agents working inside your existing site or CMS through proper permissions instead of side channels, our AI integration team builds exactly that kind of action layer.

Source: RuntimeWire