Garry Tan's $200,000 AI agent bill and the access it buys
- Garry Tan reports spending between $100,000 and $200,000 a year on AI tokens.
- Tan has published 23 Claude Code skills, called gstack, on GitHub.
- Y Combinator runs an internal agent, Stray Light, holding context on backed founders and received applications.
- Tan argues that treating agents like conventional software sacrifices much of their value.
Garry Tan, chief executive of Y Combinator, has put a figure on sustained agent use: between $100,000 and $200,000 a year in AI tokens. He gave that number in an interview with Walter Isaacson for the podcast The Builders, and two more concrete artefacts sit alongside it. Tan has published a set of 23 Claude Code skills called gstack on GitHub. Y Combinator also runs an internal agent, Stray Light, which holds what he calls a company brain covering every founder the accelerator has backed and every application it has received. TechCrunch’s account of Tan’s agent spending and the access he grants sets out the arrangements.
The spending is not the interesting part. The access model that makes it necessary is. Tan reportedly gives an agent reach into his email, his Dropbox files and his Apple Notes, including two decades of therapy notes. Stray Light, meanwhile, concentrates applicant and founder context into a single queryable system. Both arrangements trade conventional record-keeping for recall. That trade carries obligations which a demonstration of capability does not settle. An agent that can retrieve anything is also an agent whose access must be scoped, logged and revocable, and whose retention rules have to survive staff turnover.
Tan’s stated position sharpens the tension. He argues that treating agents like ordinary software, with heavy emphasis on repeatability and safety, gives up much of their value, and that personal agents will become mainstream only if costs fall sharply. That first claim deserves scrutiny rather than acceptance. Repeatability and safety are what make a system auditable. An unauditable assistant is hard to defend when it touches student records, draft manuscripts or correspondence.
The question sits close to the team’s work on honest, verifiable AI use in education and academic work. Universities hold material with statutory handling requirements, and an agent wired into email and notes inherits them. The published gstack skills are the more useful contribution here. Configuration that can be inspected and diffed is easier to govern than a closed assistant whose behaviour has to be taken on trust.
What would need to be true for this to matter beyond one executive’s setup: access limited to what a task requires, a log of what was read and changed, retention periods that match institutional policy, and a cost curve that lets ordinary staff and students run comparable tools rather than reserving them for those who can absorb six figures a year. The numbers are a useful marker, but the governance is the open question.