Peter Steinberger on what happens when millions let an agent cook

Peter Steinberger frames the talk as a direct answer to the questions that followed the explosive growth of his agent project. The origin was unusually small: while juggling agents and getting food, he wanted to send a prompt from his phone to a computer. A quick WhatsApp relay made the interaction feel less like a terminal and more like a proactive companion. Repeated attempts to show that feeling to friends produced strong reactions, and an early Discord experiment let a wider group see and test the system in public.

That momentum came with a price. The project went viral, his inbox filled with reporters, contributors, security people, and curious users, and the flood of attention made him consider deleting the whole thing. His account is not a simple celebration of growth; it is also about the emotional and operational burden of maintaining a fast-moving open-source project.

The practical lesson is that capable agents need more than a clever prompt. Steinberger returns to loops, context, verification, safety, and risk management. Long-running systems can be useful when they reduce user complexity, but they must also be designed to check their work and to operate with appropriate boundaries. In his own setup, remote machines and dedicated computers let agents run without taking over his primary laptop.

For founders, he recommends building something they genuinely want to use, developing visibility because attention is scarce, and considering hard, boring categories where a working solution earns appreciation. His final request is concrete: reliable, fast, and affordable testing infrastructure for Mac development remains harder to find than comparable Linux options.