How to Build a Self-Improving Company with AI

The speaker compares traditional companies to Roman legions, where people pass information up and down a hierarchy. They argue AI makes it possible to rethink that model by turning a company’s internal knowledge into context, skills, and software that AI can use.

A proposed AI loop includes sensors such as support tickets and product telemetry, policies that set limits, tools the AI can call, quality checks, and a learning mechanism. At YC, a monitoring agent reviews employee queries, identifies failures, and can update tools or skills; the speaker says it now writes and merges fixes overnight.

The speaker suggests using agents to find sales funnel friction, run A/B tests, and deploy the better version repeatedly. Customer suggestions could also be triaged against the roadmap, with matching requests turned into code and shipped to customers.

The speaker argues that emails, messages, and conversations need to be recorded and summarized so AI can use them. YC used about 2,000 hours of recorded office hours to generate a 150-page founder manual, which can be refreshed monthly and supplied as context to an agent.

The speaker says company data, business context, and skills are more valuable than the internal software built on top of them, which can be regenerated as models improve. They see humans as most useful at the boundary with the real world, especially in novel, ethical, high-stakes, or emotional situations, and urge founders to design for this structure from the start.