Y Combinator Chief Says He Pays Up To $200,000 A Year To Power Personal AI Agents

Y Combinator's Garry Tan told a podcast audience that he pours roughly one to two hundred thousand dollars annually into the compute credits that run his fleet of AI agents, which he relies on for both work and private life. The disclosure came on Walter Isaacson's The Builders, where Tan described how automated assistants help him with research, software tasks and everyday administration.

Tan said the agents are woven into his workflow: they assist in preparing for interviews, manage coding chores and sift through years of his personal documents stored across cloud services. He has published a collection of about two dozen agent "skills" for Anthropic's Claude Code on GitHub, hoping others can learn how to deploy similar helpers. According to Tan, the assistants don't just follow commands — they surface suggestions, anticipate follow-ups and can act with a degree of independence that he compares to a constantly present aide.

At Y Combinator, those systems are being used on a larger scale. Tan described an internal tool, dubbed Stray Light, that combines a conversational persona with a company-wide knowledge base. The agent, he said, holds searchable context on past founders, applicants and correspondence, and is being used to reshape how the accelerator supports startups and handles inquiries.

Tan also warned against treating these agents like conventional, predictable software. He argued that overemphasizing strict reproducibility and rigid controls risks throwing away the qualities that make generative assistants useful — their ability to respond flexibly and to act in ways that resemble human judgment. He acknowledged, though, that this flexibility comes at a price: token costs are high, and he called his own spending a "privilege" that most people can't afford.

That trade-off raises two questions for the wider public: who pays for broad use of such agents, and who controls them? Tan said he's skeptical of one-size-fits-all, free agents run by major platforms, and prefers the idea of personal assistants that users program and control themselves. He argued that ceding algorithmic control to a single dominant firm creates unwanted dependence.

Privacy and equity experts say Tan's setup illustrates both the promise and the risks of highly autonomous assistants. A privacy researcher familiar with the technology noted that giving an agent access to decades of private notes and email requires strong safeguards, and a policy analyst pointed out that the current cost structure will likely limit early use to affluent individuals and organizations. For those reasons, analysts say, mainstream adoption will hinge on tighter privacy protections and a big drop in operating costs.