Y Combinator’s Garry Tan uses AI agents so much, he spends up to $200K a year on tokens

Y Combinator CEO Garry Tan told Walter Isaacson on the podcast The Builders, in an episode scheduled for release on Tuesday, that he spends somewhere in the range of $100,000 to $200,000 a year on tokens for AI agents. He described the outlay as a personal passion rather than a budget line, saying he sometimes barely sleeps because he is so enthusiastic about working with his agents, and he has previously joked about suffering from what he called "cyber psychosis" in the same vein.
Tan's usage spans ordinary professional tasks and much more. He said he leans on an agent the way many people do — to research Isaacson's podcast and prepare for likely questions. Beyond that, he says the agent helps him run his life, having read all of his email, his Dropbox files and his Apple Notes, including therapy notes going back 20 years. That description is worth pausing on: a single assistant with access to two decades of personal records is a very different deployment from a coding helper, and it depends entirely on broad permission grants to the underlying data stores.
His framing for what such an agent is like is a companion that knows you and acts on that knowledge. He compared it to a Jiminy Cricket on your shoulder, running ahead of you to raise options and help you see around corners. That is a largely autonomous role, with the agent initiating suggestions rather than waiting for a prompt.
On the tooling side, Tan has published a set of 23 Claude Code skills on GitHub under the name gstack, describing the release as an effort to teach others what agents can do. Claude Code skills are reusable instruction bundles — often files plus scripts — that tell an agent how to perform a particular job, so gstack amounts to a public recipe book drawn from his own daily practice. He also identifies himself as an early adopter of OpenClaw who now uses it heavily.
One lesson he emphasized is that agents should not be handled like conventional software. He argued that some people worry so much about repeatability and safety that they give up the qualities that make agent work valuable, which he said can feel remarkably similar to collaborating with a person. This is a real design tension rather than a slogan: deterministic, tightly sandboxed workflows are easier to audit and reproduce, while broad-context, loosely supervised agents are more useful and much harder to reason about when something goes wrong. Tan's position clearly favors the second approach.
The pattern extends to Y Combinator itself. Tan said the accelerator has an internal agent, named Stray Light, which he described as having a personality along with a "company brain" holding the total context of every founder YC has backed, everyone who has ever emailed the firm, and all applications. That is a substantial corpus of sensitive material — applications in particular contain personal and commercial details founders shared under an expectation of confidentiality. How such a system is access-controlled, logged and corrected is the kind of question any organization building an equivalent will have to answer.
Tan does not think personal agents are ready for everyone yet, and he named cost as the blocker. He said the technology is too expensive right now and acknowledged that the sum he spends is a privilege most people cannot match. The implication is straightforward: at current token prices, an assistant with the depth of context he describes is a high-income tool, and mainstream adoption depends on prices falling substantially. This suggests the economics, not model capability, may set the pace of adoption for always-on personal agents.
He also sketched a preference about who should own such agents. Discussing Meta's role as a social platform controlling an algorithm many people wish it did not, he suggested imagining a personal AI that you control and program yourself. The alternative he implies is an assistant provided by a large platform, which brings convenience but puts the context and the steering in someone else's hands. For developers, that maps onto a familiar architectural choice: local, user-controlled agents versus hosted assistants run by a vendor.
Why it matters: Tan's figures give a concrete, unusually high-end data point on what heavy agent use costs today, which matters to teams budgeting for coding agents and to product managers weighing per-seat or usage-based pricing. His gstack release and the Stray Light example show how organizations are pushing agents from code assistance into company-wide context, a shift that raises real questions about data governance and permissions. And his admission that cost is the main barrier suggests that cheaper inference, rather than new capabilities, is the unlock for personal agents reaching ordinary users — an inference from his remarks rather than a prediction he made.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Y Combinator Garry Tan has jokingly admitted that he's suffered from "cyber psychosis” and barely sleeps sometimes because he’s so excited to be working with his AI agents. On a new episode of Walter Isaacson’s podcast, The Builders, Tan put a price tag on his passion.