Key points
- He uses agents to process emails, notes and therapy records.
- YC employs an internal agent named Stray Light for founder context.
Tan described this expenditure as a reflection of his heavy reliance on autonomous agents for both professional tasks and daily life management.
The figures
He shared a set of 23 Claude Code skills, known as gstack, on GitHub. His personal agent has access to therapy notes spanning 20 years. Y Combinator uses an internal agent named Stray Light.
Behind the figures
These figures were disclosed by Tan during an interview with Walter Isaacson for the podcast The Builders, which was set to publish on a Tuesday. Tan characterised the cost as a privilege not available to everyone, noting that current pricing remains too high for mainstream adoption.
What changes, what does not
Tan indicated that personal AI agents will likely become mainstream only if costs decrease significantly. He noted that users often sacrifice the human-like qualities of agents by prioritising repeatability and safety, suggesting a shift in how developers approach agent design is necessary.
What builders can do with it
Developers can build agents that operate with high autonomy, accessing extensive personal data such as emails, Dropbox files and Apple Notes. Tan’s gstack project demonstrates how to structure skills for Claude Code. Builders can create agents with persistent context, like YC’s Stray Light, which retains knowledge of all backed founders and applications.
Why it matters
- The high token cost highlights a barrier to scaling personal agents for broader consumer use.
- Tan’s open-source gstack provides a concrete reference for implementing Claude Code skills.
- The use of long-term personal data in agents raises significant privacy and security considerations.
- YC’s internal use case shows agents can maintain organisation-wide context across years of data.
Topics: AI agents, Y Combinator, Claude Code, token costs, Garry Tan
Updated
Then / now / next
Tan previously joked about suffering from cyber psychosis due to his excitement for AI agents.
He has quantified his usage, revealing an annual spend of up to $200,000 on tokens.
Mainstream adoption of personal agents depends on whether token costs decrease significantly.
Story timeline
- Then Tan previously joked about suffering from cyber psychosis due to his excitement for AI agents.
- 03 Nov 2026 Y Combinator’s Garry Tan uses AI agents so much, he spends up to $200K a year on tokens
- Now He has quantified his usage, revealing an annual spend of up to $200,000 on tokens.
- Next Mainstream adoption of personal agents depends on whether token costs decrease significantly.
The strongest doubt
Counterpoint from the sources
Tan noted that current costs are too high for mainstream adoption, calling his spending level a crazy privilege not everyone can afford.
For readers
Developers
- Use Tan's gstack as a reference for building Claude Code skills.
- Consider balancing safety requirements with the autonomous capabilities of agents.
Policy
- High token costs may limit equitable access to advanced AI assistance.
- Agents accessing decades of personal notes raise data retention questions.
Investors
- YC's internal agent use suggests operational efficiency gains from AI.
- Cost reduction is a key prerequisite for mass-market agent adoption.
Try it
Clone gstack from GitHub to review Claude Code skills.
Good result: Users can inspect the 23 skills Tan uses for his own agent workflows.
Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
What data does Tan’s personal agent access?
The agent reads his emails, Dropbox files, Apple Notes and therapy notes from the past 20 years.
What is Stray Light?
Stray Light is an internal AI agent used by Y Combinator that holds context on all founders and applicants.
Connected stories
Reading receipt
- He uses agents to process emails, notes and therapy records.
- YC employs an internal agent named Stray Light for founder context.
Still open: Will token costs decrease enough to make personal agents accessible to average users?
Still open: How will privacy regulations adapt to agents accessing long-term personal records?