What Garry Tan's token bill of up to $200,000 a year tells agent builders

What happened

According to TechCrunch, Y Combinator CEO Garry Tan told Walter Isaacson on the podcast The Builders that he spends "about maybe $100[,000] to $200,000 a year on tokens". He uses agents for countless work projects and coding tasks. He has also shared a set of 23 Claude Code skills, called gstack, on GitHub, as a way to help teach others what agents can do.

Tan said his agent has read his email, his Dropbox and his Apple Notes. He described agents as help with almost anything you do day to day, done more autonomously, and compared the experience to a Jiminy Cricket on your shoulder that will "run ahead" and help you see around corners.

Y Combinator is also using an internal agent named Stray Light, which, according to Tan, has a personality and a "company brain". He said it holds the total context of every founder YC has backed, anyone who has ever emailed YC, and all the applications.

What it means for you

The headline figure is one power user's bill. Tan himself called his spending "a crazy privilege that not everyone can do". TechCrunch reports that he believes personal AI agents will become mainstream only if costs come down. His words on the present state of things were blunt: "It's too expensive right now."

Our reading: if you are a freelancer or small business owner, a job such as writing a quote or chasing an overdue invoice is narrow. It has a clear start and a clear finish.

Tan also made a point worth weighing. He said some people worry so much about repeatability and safety that they trade away the best parts of working with an agent, which he said can be much like working with a human. For work that touches customers and money, our reading is more cautious: an agent that can act should also be able to show its work, so you can see whether it did the job properly before anything goes out.

Tan also spoke about control. Discussing Meta, he said it is a huge company that controls the algorithm, and "we all wish that they didn't at some level". He added: "So you could imagine a personal AI that you control, that you program." For builders, our reading is that this points to designing the agent yourself, around your own tasks.

Where Tool & Loop fits

In our view, building a good agent is a design skill, not just a matter of buying more tokens.

That is the ground Tool & Loop covers. It teaches agents that take a job off your list: quote a customer, chase an invoice, cut a video, check their own work. If you want to build agents like that, that is what the course is for.