Screenpipe bets the next agent leap runs through your screen

AI agents are getting stuck at the same wall: they only know what users hand them. The last two years of context engineering have been a procession of workarounds: fine-tuning that locks knowledge into a checkpoint, retrieval systems that force the user to curate which documents count, tool calls that demand the user pre-declare the right skill, and agent-maintained wikis that only work if the user trusts the model to decide what is worth remembering. Each layer asks the user to do the selecting.

Screenpipe's launch is the first product that calls the selecting the bug. The YC S26 pitch, by founder Louis, stakes the bet that the next jump in agent usefulness comes from continuous, on-device capture of the screen and audio the user already produces. The agent stops asking which app, which doc, which skill: it watches the work and answers.

The trade the user is being asked to make is the actual story. A tool that records the screen all day is not a productivity upgrade; it is a new kind of memory lease, where the agent's competence and the user's exposure scale together. The local-first posture is what makes the bet legible rather than alarming, but it does not erase the question. As agents move from opt-in assistants to ambient observers, the divide between products that capture by default and products that ask permission becomes the new privacy frontier.