Autonomous companies from a single prompt

Key points

Key takeaways from a Génération Do It Yourself interview with NanoCorp founder Pierre-Louis Biojout (September 2026):

Autonomous companies from a prompt. NanoCorp lets anyone turn an idea into an autonomous business from a single prompt: a CEO agent deploys a site, sets up payments and runs operations without human intervention. Biojout reports roughly 1.7 million agent runs to date, consuming 20–50 billion tokens a week (around $25,000 a week in spend).

Profitable since summer 2026. After an earlier round from Y Combinator (W24 micro-seed), NanoCorp has been profitable since this summer, a dynamic Biojout says has changed the company as it scales on agent labour rather than human headcount.

Agents, not chatbots. An agent is an entity that observes its environment and acts within it to reach an objective: prompts become goals, and the model loops for many turns with tools (what he calls a harness) instead of handing the ball back after each answer.

Models increasingly work for other models. Agentic models now answer other agents more than humans: a tool like Claude Code internally deploys a team of ten to hundreds of agents to reach its objective, and recursive self-improvement lets an agent rewrite its own code when it errs.

Agents still need verifiers. Biojout cites a practical failure mode — an agent citing basketballer Rudy Gobert’s defensive awards as four when he has five — and argues agents must also check their facts against up-to-date sources.

OpenAI’s agent-run data centre. He points to OpenAI ’s first data centre operated 100% by agents (autonomous research), where agents get full GPU access and run training experiments — likely behind models such as GPT-6 Astra.

Two Millennium problems claimed solved. After OpenAI said it had solved two of the Millennium Prize problems (Navier–Stokes and the Hodge conjecture), Biojout spent three hours with his siblings just to grasp the surface of the Hodge problem — “we are monkeys to the AI now”.

Launch a business while doing (almost) nothing. His thesis: build by making the maximum number of non-lethal errors, and when you win, win very hard — while companies run largely by agents remove the context and workflow friction that make AI hard to slot into existing organisations.

Read more: Génération Do It Yourself on YouTube