Rabbit shows off the AI agent it should have launched with

The Humane AI Pin has collapsed, but Rabbit is still kicking. The company published a blog post and video today showing off a “generalist Android agent,” slowly controlling apps on a tablet in much the same way that Rabbit claimed its R1 device would over a year ago. (It couldn’t, and can’t.) The work builds on LAM Playground, a “generalist web agent” Rabbit launched last year.

The engineers don’t use the Rabbit R1 at all for the demonstration. Instead, they type their requests into a prompt box on a laptop, which translates them to actions on an Android tablet. They task it with things like finding a YouTube video or locating a whiskey cocktail recipe in a cocktail app, gathering the ingredients, and then adding them to a Google Keep grocery list. At one point, they ask it to download the puzzle game 2048 and figure out how to play it, which it does, albeit slowly.

The model generally does the things they ask, sometimes well and sometimes with quirks like sending a poem over WhatsApp one message at a time instead of in a single block. One of the engineers wonders if they should have asked it to use line breaks in their prompt, but they don’t go back to try again.

Rabbit’s AI agent is clearly still a work in progress, as it has been since the R1 launched with almost none of the capabilities that founder and CEO Jesse Lyu presented in January 2024. Rabbit has steadily rolled out updates, like the ability to train its AI agents to complete specific tasks or prompt it to remake its own interface. The examples it presented today are “only the core action loop an Android agent completes,” according to Rabbit’s blog post. The company promises to share more about its “upcoming cross-platform multi-agent system” in coming weeks.