AI agents are becoming part of Y Combinator CEO Garry Tan’s daily work and personal routines. He says he spends about maybe $100[,000] to $200,000 a year on tokens, describing the cost as a privilege that not everyone can afford.
Agents for work and everyday life
Tan uses agents for research, coding and other work projects. Before appearing on Walter Isaacson’s podcast The Builders, he also used an agent to research the podcast and prepare for possible questions.
His use goes beyond individual tasks. Tan said the agent has read his email, Dropbox and Apple Notes, including therapy notes going back 20 years. He imagines an agent as a helper that knows its user and can take initiative, suggesting what to consider or helping them anticipate what may come next.
That vision depends on an unusually broad view of personal information. An agent with access to a person’s notes and messages may be able to offer context across daily activities, rather than responding only to a single prompt. Tan’s account illustrates both the appeal of that arrangement and how much access it can involve.
A company agent with institutional context
Y Combinator is also using an internal agent called Stray Light. Tan said it has a personality and a “company brain,” with context about every founder the accelerator has backed, people who have emailed the organization and its applications.
He described the agent as part of “rebuilding the way we operate at YC.” The example shows how the same idea can extend from an individual assistant to an organization: an agent draws on accumulated information to help with work across the institution.
Tan has shared a set of 23 Claude Code skills called gstack on GitHub to help teach others what agents can do. He is also an early OpenClaw adopter and now a power user.
Why Tan says agents need a different mindset
Tan argues that people should not treat AI agents exactly like traditional software. He said a focus on repeatability and safety can lead people to trade away some of the qualities that make an agent useful, including its ability to work in ways that can feel similar to a human.
His comments point to a tension in agent design. The more an agent can adapt to context and act with some autonomy, the more users may value that flexibility. At the same time, Tan’s examples involve access to extensive personal or organizational records, making trust and control central to whether people will want such help.
Cost and control will shape wider adoption
Tan believes personal AI agents could eventually become mainstream, but says the price must come down. He called the amount he spends “a crazy privilege that not everyone can do.” For now, his experience reflects a version of agent use that may be out of reach for many people.
He also questioned whether free agents controlled by large technology companies are the right answer. Discussing the influence of Meta as a social media company, he said people may wish they had more say over the algorithm. He described a future with a personal AI that users control and program themselves.
That future, as Tan describes it, rests on two conditions: agents must become more affordable, and users must have meaningful control over them. His own use shows what an agent can do when it has broad context and handles a range of tasks. Whether that model spreads will depend on making its benefits accessible without taking control away from the people who rely on it.