AI agents come with a price tag, and now one of the best-known names in the startup world has decided to share how much he pays for them. According to TechCrunch, Garry Tan, CEO of Y Combinator, said he spends somewhere between $100,000 and $200,000 a year on tokens alone to keep his agents working. He made the comment on the podcast The Builders, hosted by writer Walter Isaacson, in an episode set to premiere this Tuesday.
The number grabs attention, of course. But the more useful takeaway is something else. It gives a sense of how much it can cost to use agents very heavily, both at work and in personal life. For founders and people working in marketing and tech in Brazil, this math helps when thinking about cost and AI return on investment before handing everything over to an agent.
How Tan uses his agents
According to the TechCrunch report, Tan uses agents across a bunch of professional projects and coding tasks. For the interview itself, for example, he had his agent research the Isaacson podcast and prepare him for the kinds of questions that might come up. So far, nothing very different from what plenty of people already do with AI every day.
The difference is on the personal side.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
According to the outlet, the agent has already read his emails, his Dropbox files and his Apple Notes, including therapy notes collected over about 20 years. He compares the tool to a kind of Jiminy Cricket always sitting on his shoulder: someone who knows him, helps with almost everything and sometimes jumps in with suggestions he had not even considered yet. He is so excited about it that, as TechCrunch notes, Tan has joked in the past that he suffers from a sort of cyber psychosis and that he sometimes barely sleeps because he is so eager to work with his agents.
He also decided to share some of what he learned. Tan published gstack on GitHub, a set of 23 skills for Claude Code, meant to show other people what agents can do. And he is not the only one using them. At Y Combinator, TechCrunch reports, there is an internal agent called Stray Light, which has something like a personality and a kind of company memory, holding context on every founder the accelerator has backed, every application it has received and everyone who has ever emailed it. Tan said this agent is helping rebuild the way YC operates.
Agents are not traditional software, Tan says
Tan was one of the early adopters of OpenClaw and now describes himself as a power user. One lesson he took from that experience, according to the report, is not to treat agents like ordinary programs. In his view, some people worry so much about predictability and safety that they end up giving up exactly what is best about working with an agent, which can behave a lot like a person.
Worth noting: this is how Tan sees it, not a consensus. In companies, security and predictability are still serious concerns, as we will see further down.
Even Tan thinks it is too expensive
As big a tech enthusiast as he is, Tan admits cost is a major hurdle. According to TechCrunch, he said prices are too high right now and acknowledged that spending this much on agents is a privilege few people have. In his view, personal agents should go mainstream, but that depends on prices coming down.
Still, for him, the answer is not simply waiting for free agents offered by big tech companies. Pointing to Meta as an example of a company that controls the algorithm everyone consumes, Tan argued, according to the outlet, for the idea of a personal AI that users control and program themselves.
What this means for Brazilian businesses
Nobody needs to spend what Tan spends to get value from agents. Even so, his case offers a clear clue: AI token spending tends to grow as an agent gains autonomy and receives more context. The more it reads, remembers and executes, the bigger the bill. That is why the cost of AI agents needs to be part of planning from day one, instead of showing up as a surprise on the invoice. Some companies are already tackling this, as shown by how Okta uses MCP to cut token costs in AI agents.
When thinking about return, the most useful question is simple: what task does the agent handle, and how much time or money does that save? Picture a small agency that sets up an agent just to organize briefs and summarize meetings. Measuring the gain is much easier than if it tried, right out of the gate, to build an agent that does everything. Starting with a well-defined use case is usually the safest path, as we explain in AI agents for small businesses.
Another point deserves attention. The Tan example shows how much an agent can see when it gets broad access: emails, files, even very private notes. In a company, that calls for clear rules on what data the agent can access and who is accountable for what it does. It pays to learn best practices to test and audit agents safely before handing over sensitive information, especially customer data.
What Tan said works as a snapshot of where agents can go when money is no object. For most businesses, though, the math is different. Choosing the right task, tracking token usage and measuring results tends to pay off more than trying to keep pace with those at the top of Silicon Valley.
