Tech • AI • Robotics • Game
He started a business without doing anything (or almost) - Pierre-Louis Biojout
7/10TL;DR
Pierre-Louis Bijou, founder of Nanocorp, argues that AI agents are rapidly shifting from chat assistants to autonomous workers, reshaping startups, software and white-collar work.
KEY POINTS
From chatbot to agent
An AI agent is defined as an entity that observes its environment and acts to achieve a goal. The change from earlier chatbots is that the system no longer waits for a human after each answer: it can loop through tasks, call tools, browse data and execute multi-step objectives. A request such as building a pitch deck can now involve market research, drafting, revisions and delivery without constant supervision.
Agents still fail on basic reasoning
Despite progress, current systems remain uneven. Some agents can search the web, write code and automate workflows, yet still miss deductions that a teenager would make. Their biggest weakness often appears when relevant context is not digitized and remains in documents, fragmented systems or the head of a human employee.
Nanocorp’s operating scale
Nanocorp, launched in March 2026, aims to let users turn an idea into an autonomous business from a prompt. The company says it has already run agents about 1.7 million times and is now profitable after earlier fundraising. Internal spending on model usage reaches roughly 20 billion to 50 billion tokens per week, for a weekly bill that can hit about $25,000.
Recursive improvement is becoming practical
Beyond executing tasks, some agents are now designed to modify and improve their own code. This form of recursive self-improvement allows software to react to errors, optimize its workflows and iterate without waiting for a developer to rewrite each component. That approach is already being used in production systems, though it remains imperfect.
A new startup model with fewer humans
Bijou describes a startup model centered on agents rather than large teams. In that framework, founders rely on software workers that can prospect leads, prepare documents, manage pipelines and coordinate operations. The logic is economic as much as technical: software scales faster than human hiring, and a company with minimal headcount can move unusually quickly.
The frontier is automated research
The next step is not just replacing office tasks but automating research itself. Large AI labs are already using agents to test training ideas, run experiments on GPU clusters and evaluate model performance, reducing the need for human researchers in some loops. Bijou points to data centers increasingly operated by agents and to advanced systems that may already be contributing to frontier model development.
Intelligence may outrun explanation
One consequence of that progress is that AI can solve or analyze problems that many educated users cannot fully understand. Bijou cites recent claims that advanced systems tackled major mathematical problems such as Navier-Stokes and the Hodge conjecture, while humans may need hours just to grasp the statement of the problem. That creates a paradox: machines may produce correct answers before they can explain them in a way ordinary users can follow.
French elites are seen as too risk-averse
Bijou, a Polytechnique graduate who finished in 2023, criticizes the career conservatism he sees among many graduates of elite French schools. In his view, too many still prioritize stable posts in large groups or the civil service rather than entrepreneurship. He argues that this mindset clashes with the current AI cycle, where speed, experimentation and tolerance for non-fatal mistakes matter more than polished credentials.
Europe is losing ground in foundation models
On the competitive landscape, Bijou sees no credible third bloc between the United States and China in top-tier AI models. He says Mistral once looked competitive but now appears to have shifted toward services, deployment and consulting rather than trying to win the frontier model race. That reflects a broader view that model quality will keep improving while costs fall, concentrating power around those best able to exploit the technology.
Toward firms with zero employees
The most radical projection is an economy where many companies operate with no human staff at all. Bijou describes that as a possible end state because firms built entirely from software agents scale more easily than organizations tied to one person per entity. He also expects legal systems to explore new structures for agent-run organizations, with places such as Delaware and Argentina already seen as potential testing grounds.
CONCLUSION
The central bet behind Nanocorp and similar ventures is that autonomous agents will become a basic economic resource, not just a productivity tool. If that happens, the main divide will be between organizations that redesign themselves around AI and those that continue operating at human speed.
Ask a question