Garry Tan says founders should spend heavily on AI agents

What’s the deal? In an episode of the a16z podcast released in August 2026, Garry Tan, CEO of Y Combinator, argued that founders should give themselves permission to spend heavily on AI-agent usage, even when the token bill is unusually high. He described loading as many as one million tokens, or roughly 800,000 tokens, into an agent context so that a founder can access a level of capability he characterised as “living in 2028”.

Why it matters: Tan estimated that using agents at full strength could cost $50,000–$100,000 a year, but said the investment can make sense for a CEO or founder if it accelerates experimentation and execution. His recommended workflow is to turn successful agent processes into reusable instructions — to “skillify” them — so that a task can be repeated consistently rather than rebuilt from scratch. He described a markdown file containing those instructions as “an employee” that can perform the job repeatedly. The comments place Tan among the more vocal proponents of “tokenmaxxing”, the practice of allowing AI agents to consume large token budgets in pursuit of stronger results, while other technology leaders have argued that output and efficiency matter more than raw token consumption.

The debate reflects a broader question for startup operators: whether expensive frontier-model usage should be treated as discretionary tooling or as a form of leverage for small teams. Tan’s position is that founders are unusually well placed to absorb the cost when the work can be converted into durable processes and repeated across the organisation.

Read more: Business Insider

Image Credit: Initialized Capital