Garry Tan's AI Agent Bill: What a $200K Token Habit Says About Agent Costs

What Garry Tan said

TechCrunch's Julie Bort reported on November 2, 2026 on an interview Y Combinator's chief executive gave to Walter Isaacson for the podcast The Builders. The headline number was Tan's own rough estimate of his token spend: "about maybe $100[,000] to $200,000 a year on tokens," which he called "a crazy privilege that not everyone can do."

The details behind the number matter more than the number. According to TechCrunch, Tan uses agents for work projects and coding, used one to research and prepare for the interview itself, and runs parts of his life through one that has read his email, his Dropbox, his Apple Notes, and therapy notes going back 20 years. He described the effect as a "Jiminy Cricket on your shoulder" that suggests things before he asks. He has published gstack, a set of 23 Claude Code skills, on GitHub, and TechCrunch describes him as an OpenClaw power user.

Inside YC, an agent called Stray Light "has a personality" and a "company brain," with context on every founder YC has backed, everyone who has emailed YC, and all applications. Tan said YC is "rebuilding the way we operate." His view on the mass market was blunt: personal agents go mainstream only if costs fall. "It's too expensive right now."

Why one person's agents can cost that much

At the top of his range, the bill works out to roughly $550 a day. That is not one big model call. It is the shape of agent work, multiplied.

• Context is re-read, not remembered. A model has no memory between calls. An agent that "knows" years of email and notes either sends large slices of that material with every request, or searches it. Sending it is the expensive path, and it repeats on every step of every task.
• Agents loop. One request becomes plan, search, read, draft, check, revise. Each step resends the growing transcript, so input tokens grow faster than the work does.
• Proactive means always running. An agent that suggests things before it is asked has to look at new information continuously. Background monitoring is billed whether or not it finds anything useful.
• Frontier models for everything. Heavy users default to the strongest model because it fails less, and pay frontier rates for steps a smaller model could handle.

None of this is unique to Tan. It is the default cost curve of any agent with broad context and autonomy, which is exactly why he expects cost to gate adoption.

A company brain is a retrieval problem

Stray Light's description, context on every founder, every email, every application, is the clearest example of where agent costs actually live. No context window holds an accelerator's entire history, and resending even a fraction of it per question would be ruinous. The workable design is retrieval: index the corpus once, fetch the handful of passages relevant to the current question, and let the model reason over those.

Retrieval also changes what happens on the second ask. Many organisational questions repeat: what did this founder say last batch, which companies worked on this problem, what did partners decide about a similar case. If the answer, or the evidence for it, is already indexed, the agent retrieves it instead of regenerating it from raw context. On repeated or reference-heavy work like this, RDK's stacked retrieval cuts token spend 80 to 90 percent, because the private vault answers 40 to 65 percent of queries, the public network 15 to 20 percent, and the LLM is the fallback for only 5 to 10 percent.

The cost levers, in order

For anyone whose agent bill is heading in Tan's direction, the levers are well understood:

Repeatability versus personality

TechCrunch reported Tan warning against treating agents like traditional software, saying some people worry so much about "repeatability and safety" that they lose what makes agents feel human. That is a real trade-off, not a slogan. A personal agent drafting suggestions can afford variety. An agent with access to someone's email, therapy notes, and an organisation's applicant records needs firm boundaries on what it may read, what it may send, and where that data is stored. The practical answer is to keep the creative layer loose and the data layer strict: private, access-controlled stores for sensitive material, and explicit approval for anything that leaves the system.