Builder.io open-sources Agent-Native: agents share actions and state with the UI

Rajashekar Sudireddy @raja_sudireddyOriginally posted 2026.10.25

Open original on X

http://Builder.io just launched Agent-Native in beta, an open-source TypeScript framework for building AI agents that come with their own app. The idea cuts against how almost every agent ships today. The agent lives in a chat box. You can't see what it's doing except in words, and fixing one detail means writing another prompt, waiting for the agent, paying for the tokens, and checking nothing else broke. Builder's blog puts it bluntly: fixing one color shouldn't mean writing a prompt, waiting for the agent, paying for the tokens, and checking that nothing else changed. Agent-Native connects the agent and the UI through three shared foundations. Shared actions: define a capability once, and the agent calls it as a tool while the UI calls it from code, both through the same validation, permissions, and implementation. Shared data keeps their work in sync. Shared application state gives the agent context: it knows what's on your screen, the current page, the selected record, the active view, so it understands what you mean by "this." That stands in contrast to computer-use agents, which work through the UI the way a person would: look at the screen, decide where to click, act, check the result, one step at a time. Builder calls that slow and token-heavy. Agent-Native agents skip the screen and call the same actions the UI uses, so they finish faster and spend fewer tokens. With no screen to depend on, they can also run from Slack, on a schedule, or at another agent's request. One defined capability also gets exposed across six surfaces: UI, agent, HTTP API, CLI, MCP, and A2A. Every Agent-Native app is an MCP server, so Claude, ChatGPT, Cursor, and Codex can call your app's actions directly. And through the A2A protocol, one agent can hand a task to another, which handles it with its own tools and returns the result. The team features matter as much as the architecture: built-in auth, permissions, organizations and teams, plus provenance. You can require approval before sensitive actions and always tell whether a change came from you, a teammate, or the agent. The thesis is simple. A chat box gives the agent freedom but takes away control. The agentic apps that win will pair the freedom of a prompt with the control of an app. Clips, Slides, and Design ship as open-source examples built on the framework, so the whole pattern is inspectable. Worth testing on a real workflow before you take the architecture on faith.