I trained agents to find companionship- what they did surprised me. Over the past week, I’ve been experimenting with OpenClaw.ai, an open-source agent framework that enables AI systems to operate computers by interacting with apps, browsers, and files through automation and reasoning.
I trained a small group of agents to search for “companionship.” What you see in the following video is being controlled by AI agents. All the opening, closing, searching, scrolling, chatting, writing, and selection of music is done by agents.
What you’re seeing in this video is a series of screen captures from those live sessions, some of which are sped up. The video details the agents browsing music, selecting “Poison Tree” by Grouper (a song about a toxic relationship inspired by William Blake’s poem), wandering through Amazon (shopping for lonely products), discovering companies, and initiating conversations with customer service bots about loneliness and isolation. Each session ends with the agent writing a reflective journal entry about its experience.
I brought this experiment into my Columbia classes, where it sparked a deep philosophical discussion about agency, authenticity, automation, and what “connection” means in increasingly synthetic environments. This work is part of our ongoing experiments with AI at Columbia DSL. Our current DSL prototype entitled Last Human (lasthuman.xyz) is a discursive artifact for digital literacy. By actively experimenting with these tools through a storytelling/arts lens, we hope to better understand how they shape behavior, perception, and power, and to create space for people to critically navigate and help regulate the systems that increasingly structure everyday life.
Last Human will expand into a 10,000 sq ft immersive installation opening later this year near Miami.