Y Combinator’s CEO, Garry Tan, uses AI agents so intensively that he admits to spending between $100,000 and $200,000 annually just on the tokens needed to power them. That’s part of a broader setup where he leans on digital helpers not only for work but to manage almost every corner of his personal life.
Living with AI agents on and off the clock
Tan treats AI agents as more than tools. They’ve been granted sweeping access—reading his email, Dropbox files, Apple Notes, and even therapy notes going back two decades. He regularly taps them for research, preparing for interviews, managing day-to-day tasks, and even making suggestions or anticipating needs, describing them as “autonomous helpers” who know him well and proactively support his life.
At Y Combinator, the use of agents goes deep. Tan says there’s an internal AI agent called “Stray Light” that has personality and memory—it contains context about every founder YC has backed, plus data on every application they receive. Another initiative from Tan is “gstack,” a public repository with 23 Claude Code skills, meant to illustrate what agents are capable of doing.
Cost, control, and the future of personal AI agents
When it comes to the financial cost, Tan acknowledges this level of reliance isn’t feasible for most people. The token costs associated with powering numerous agents add up fast, and that price tag, while justified in his view, is out of reach for everyday users.
But he also warns against relying on generic agents delivered by large tech companies. For Tan, the ideal shape of personal AI agents involves individual control and programmability—rather than being subordinate to some big company’s algorithm. He believes agents must become more affordable, more private, and more under user control if they’re going to gain mainstream adoption.
This system demands trade-offs. Tan says some people sacrifice too much in safety and predictability when they use agents—losing many of the parts that make them uniquely humanlike. He treats agents more like collaborators than tools, which creates a different mindset about risk, repeatability, and efficiency.
AI agents are rapidly moving from novelty to necessity for some in tech power circles. Tan’s spend on tokens reflects not only his conviction in agents’ power but how far early adopters have already woven them into everything from inbox cleanups to strategic decisions.
What this means: Tan is showing one possible endpoint for AI in daily life. Personal agents that learn, anticipate, act—and at significant cost. If AI is ever going to reach a mass audience, it’ll need to deliver similar power without the premium price tag, without the loss of control, and without relying on large gatekeepers. Cost reduction, privacy, adjustable autonomy—these are the levers the next wave needs to pull.
