Y Combinator’s Harry Tan spends up to $200,000 a year on AI agents

An assistant with a long memory

Tan uses an agent to research and prepare for a podcast with Walter Isaacson, including anticipating possible questions. It also searches across his email, Dropbox files and Apple Notes, as well as records from therapy sessions over the past 20 years.

He describes agents as helpers for everyday tasks, able to take on some work independently. His analogy is Jiminy Cricket: a companion on your shoulder that knows you, offers help and can suggest what to do or flag what may be coming.

Tan began using OpenClaw early and now works with it extensively. He argues that agents should not be treated like ordinary software. In his view, systems designed too narrowly around repeatability and safety can miss the potential of agents that interact with people almost like other people.

Y Combinator’s internal agent

Y Combinator also uses an agent called Stray Light to change how the company works internally. It has a personality and access to a shared knowledge base about the organization.

According to Tan, its context includes every founder Y Combinator has supported, everyone who has ever written to the company, and all its applications. That makes the agent more than a personal assistant: it can draw on an institutional memory built from the company’s relationships and work.

The cost of personal AI

Tan sees cost as the main obstacle to personal agents becoming commonplace. His own spending, he acknowledges, is a privilege most people cannot afford.

The alternative — free agents controlled by major technology companies — raises a different concern. Discussing Meta’s influence on society, Tan said the company controls the algorithm, and many people would prefer that it did not. He sees room for personal AI that people control and program themselves.

My read is that the $200,000 figure is not just a price tag; it marks the gap between an agent that can build a detailed picture of one person and a tool ordinary users can realistically keep running. The unanswered question is how that kind of personal control can become affordable without handing the underlying system to a company whose decisions users cannot shape.