Y Combinator president and CEO Garry Tan estimates that he spends roughly $100,000-$200,000 a year on AI tokens for agents used in research, coding, work projects and personal tasks. The usage credits cover prompts, generated code and agent activity on large language models, including Anthropic's Claude Code. Tan has described the approach as "tokenmaxxing," arguing that heavy AI spending can let a small team or individual produce work associated with a much larger organization. He has shared gstack, a package of 23 Claude Code skills designed to simulate an engineering team, and the GitHub repository has attracted between 100,000 and 136,000 stars. Tan says Claude Code agents built his Garry's List blogging platform in about five days for approximately $200 in tokens. He has also described agents as systems capable of substantial autonomy, access to extensive personal information and task anticipation, while warning that costs must decline before they become mainstream. Y Combinator uses an internal agent called Stray Light, which Tan said has a personality and a company brain containing context on founders, contacts and applications. His heavy reliance on Anthropic's tools also highlights the exposure companies face to model-provider pricing, availability and policies.