I came across this YC talk recently and thought it was worth sharing.
The speaker is YC General Partner Tom Blomfield. His argument: companies should not just become AI-native. They should become self-improving.
Talk: How to Build a Self-Improving Company with AI
Most AI conversations stop at productivity. Copilot, code faster, ship more. Tom asks a different question: what if the company itself could improve while you sleep?
Tom points out that most companies still run like Roman legions: information flows up through layers of management, decisions flow back down. Humans are the pipes. AI breaks that model.
The real leverage is not making engineers 20% faster. It is making the company's knowledge legible to AI: emails, Slack, support tickets, meetings, product data, and the tacit know-how trapped inside people's heads.
Once that context is machine-readable, you can build recursive loops: sense, decide, act, evaluate, improve, without waiting for a human in the middle.
The YC example is what sold me. They built an agent that monitors which internal queries fail, figures out what tool or database view is missing, writes the fix, opens a merge request, gets it reviewed, and deploys it. The next morning, the same query works. No human touched it.
Tom's line that stuck with me: "burn tokens, not headcount."
For people building AI agents, I think this is the important framing shift.
The moat is probably not the model layer itself. It is the context, tooling, evals, feedback loops, and the system that keeps improving over time.