Garry Tan, CEO of Y Combinator, told Walter Isaacson on an episode of The Builders podcast that he spends “about maybe $100[,000] to $200,000 a year on tokens” for the personal AI agents he uses. The episode was scheduled to publish on Tuesday.
Tan said he relies on agents for numerous work projects and coding tasks. He has also published a set of 23 Claude Code skills called gstack on GitHub, which he described as a way to teach others what agents can do.
From research assistance to running his life
Tan described two main uses for his agents: they help him prepare for interviews and research topics — for example, the podcast with Isaacson — and they also assist with personal matters. “It’s read all of my email and all of my Dropbox. It’s read, like, all of my Apple Notes, even my therapy notes going back 20 years,” he said.
He compared an AI agent to having a “Jiminy Cricket on your shoulder,” able to act somewhat autonomously, suggest actions, and help anticipate needs by "seeing around corners." The agent can proactively propose ideas or next steps rather than waiting for explicit instructions.
Don’t treat agents like traditional software
As an early OpenClaw adopter and now a power user, Tan said one lesson is not to treat AI agents as conventional software. He argued that excessive focus on “repeatability and safety” can sacrifice the interactive, human-like qualities that make agents valuable.
Y Combinator itself is using agents to change internal operations. Tan said they have an internal AI agent called Stray Light that has a personality and a “company brain.” According to him, Stray Light holds the full context of every founder Y Combinator has backed, anyone who has emailed them, and all applications.
Costs and control are barriers to mainstream adoption
Tan emphasized that a major obstacle to mainstream personal agents is cost: “It’s too expensive right now,” he said, noting that his own spending is “a crazy privilege that not everyone can do.”
He also warned against relying solely on free agents controlled by large tech companies. Discussing the kind of future he wants, and Meta’s influence as a social platform, Tan said he would prefer personal AI that users control and program rather than an algorithm controlled by a single huge company.
Why it matters
Tan’s account illustrates both the current capabilities of personal AI agents — from productivity and coding to deeply personal data handling — and the practical barriers to broader use, namely expense and governance. By publishing his gstack skills and discussing his practices publicly, he aims to demonstrate what agents can do while highlighting that lower costs and user control are key to wider adoption.
