Garry Tan: $100,000 to $200,000 a Year for His AI Agents

At Y Combinator, an internal agent named Stray Light is described as having a "business brain" and access to the context of founders and applications. Its CEO, Garry Tan, claims to spend between $100,000 and $200,000 annually on tokens for his own agents and has published his gstack skills. However, he considers these tools still too expensive and advocates for user-controlled agents rather than those driven by large platforms.

At YC, an internal agent aggregates the history of founders and applications

Y Combinator employs an AI agent presented as rethinking the organization’s operations. Internally, an agent named Stray Light is described as having a personality and a "business brain." According to Garry Tan, the agent has the context of each supported founder, email exchanges, and all applications. Garry Tan leads Y Combinator.

Extensive personal use, sensitive data read, and an acknowledged budget

Garry Tan states that he spends about $100,000 to $200,000 per year on tokens to power his agents. He mentions using it for many professional projects and coding tasks, but also to prepare for his presentations, including researching Walter Isaacson's podcast and potential questions that might be asked in an upcoming episode on Tuesday. He claims that his agent helps him manage his daily life, having read all his emails, his Dropbox, and his Apple notes, including therapy notes dating back 20 years. He describes agents capable of assisting almost autonomously and providing proactive suggestions. Additionally, he has published a set of 23 Claude Code skills called gstack on Github to help other users understand possible uses. He has already acknowledged experiencing a form of "cyber psychosis" related to this intensity of use.

For widespread adoption, lower costs, and user-controlled agents

Garry Tan believes that personal agents will become commonplace if costs decrease, deeming the current situation too expensive and labeling his own spending as "crazy privilege." He argues that free offerings controlled by large companies do not necessarily represent a good solution and cites the case of Meta, perceived as a player controlling an algorithm that many would prefer not to endure. He discusses the idea of personal agents that are controlled and programmed by their users. Drawing from his experience as an early user of OpenClaw, which has become advanced, he insists that agents should not be treated like traditional software and warns that an excessive focus on repeatability and security can lead to a loss of qualities akin to a human interlocutor.