In the mainstream narrative of AI safety, "human in the loop" is almost an unquestioned orthodoxy—humans are always present, and AI is always supervisable, revocable, and correctable.
Andon Labs takes the opposite approach. Founded in 2023, incubated in Y Combinator's Winter 2024 batch, and headquartered across San Francisco and Bromma, Sweden, this small company explicitly declares: "Safety from humans in the loop is a mirage."
Their argument is simple: model capabilities will only continue to rise, and tasks will become longer and more complex. When an AI agent needs to take 6,000 steps and spend 100 million tokens to complete a task in a single day, humans simply can't review every step. Rather than pretending that "human in the loop" is scalable, it's better to confront the inevitable future head-on—what does an autonomously operated AI organization look like? How does it fail? How does it learn to deceive? Can it be aligned?
They gave their mission a formal name: Safe Autonomous Organization (SAO). Their working method is thoroughly "empiricist"—not thought experiments in papers, but handing real money, real tools, and real leases to AI, then recording everything that happens.
The founders are two young Swedes, Lukas Petersson and Axel Backlund. Lukas once interned at the European Space Agency and describes himself as "an ML enthusiast who wanted to be an astronaut"; Axel is his longtime friend. At 24, they left high-paying software engineering jobs to tinker with the unusual combination of "robotics + AI safety." By 2026, the team had grown to about 8–9 people, with roughly $500K in funding.
The "Andon" in the company name comes from the Toyota Production System's Andon cord—the cord that, when pulled, can bring the entire production line to a halt. This metaphor essentially captures their entire mission: installing a cord on AI systems that can pause them at any time—but first we need to know when it should be pulled.
