How to Build a Self-Improving Company with AI

In plain English

The speaker says most companies are built like a Roman army: layers of managers passing information up and down. He argues AI makes that structure unnecessary. Instead of giving staff a tool to be 20% faster, run parts of the company as loops.

The system senses what goes wrong, applies rules, uses tools, checks its own quality and learns. His example is an internal assistant at Y Combinator. A second AI watches where it fails, and overnight it writes and ships a fix, so the same question works the next day.

Three ideas worth keeping

• Record and organise what your company knows, because AI can only use what has been written down.
• Treat the software on top as disposable, but guard the data and know-how, which are the valuable part.
• People still matter at the edges: high-stakes, emotional or novel situations, and sales conversations.

What it means for your business

Start by capturing how your business actually works: emails, calls, procedures, decisions. Then pick one repeating process, like handling customer queries, and see whether it could improve itself with a person supervising. He admits nobody has fully done this across a whole company yet, so treat it as direction, not a finished recipe.

What it means for software teams

Give the AI a clear set of tools, rules about what needs human sign-off, and automatic checks before anything ships. Let a monitoring agent propose fixes from real failures. He also says simple internal dashboards can now be generated on demand and regenerated as models improve.

What we found, and what we did

Our improvement loops are partly automatic and partly human. Automatically, a brake pauses automatic merging if too many recent changes have to be undone, and a weekly job re-measures how fast our automated pipeline runs and how much it needs people, with an alert if that report stops arriving.

By hand, after a bad week we go through the agents’ own session records for patterns of failure and turn them into fixes, written rules for the pipeline and guidelines for our AI reviewer.

Over the same period the pipeline has grown from about 20 changes a day in September to about 45 a day now, with people approving under 10% of them instead of about 70%.

We would not call it a fully self-improving company; it is closer to the Plan, Delegate, Assess, Codify cycle in our field note Compounding Engineering: The 4-Step Loop.

Worth watching

• 02:47: The five layers of an AI loop, from sensing to learning.
• 04:08: The overnight fix example at Y Combinator.
• 08:12: Why recording everything makes a company readable to AI.