AI agents are not cheap when you use them all day. Y Combinator CEO Garry Tan just put a number on it. He says he spends about $100,000 to $200,000 a year on tokens to run his agents.

I'm Nitin, and I coordinate projects at Debugged Pro. Our team uses AI tools on client work every week, so I pay close attention to what they really cost. Tan's number is extreme, but it raises questions every agency and business in India and the USA should ask. How much should you spend, what should you let an agent touch, and who pays?

Y Combinator CEO Garry Tan speaking on stage at SXSW, where he has talked about using AI agents — Garry Tan at SXSW. Image: Getty Images via TechCrunch

What Garry Tan Said About His AI Agents

The comments come from a new episode of Walter Isaacson's podcast, The Builders. TechCrunch reported the interview on November 2, 2026, in a story by venture editor Julie Bort. The episode was set to publish on Tuesday.

Here are the key facts from the report:

The spend: "I spend about maybe $100[,000] to $200,000 a year on tokens," Tan told Isaacson.
The work: He uses agents for many work projects and coding tasks. He also uses one to research and prep for interviews like this one.
The access: His agent has read his email, his Dropbox and his Apple Notes. He said that includes therapy notes going back 20 years.
The tooling: He has shared 23 Claude Code skills, called gstack, for free on GitHub.
The caveat: He said the cost is "too expensive right now" and called his spending "a crazy privilege that not everyone can do."

Tan is also an early OpenClaw adopter and now a power user. He believes personal AI agents will go mainstream, but only once costs come down.

YC's own internal agent

Tan also said Y Combinator uses an internal agent named Stray Light. He described it as having a personality and a "company brain." According to Tan, it has the full context of every founder YC has backed, everyone who has emailed YC and all the applications.

That is a big deal. It shows a well-known firm moving agents from side projects into core operations. It also shows how much sensitive data these systems can end up holding.

What $200K a Year on AI Agents Really Looks Like

Big yearly numbers are hard to picture. So we broke Tan's range down into smaller pieces using simple division.

Chart breaking down Garry Tan's AI agents token spend of $100K to $200K a year into monthly and daily costs — Tan's yearly range split by month and by working day. Data: TechCrunch, Nov 2, 2026. Chart: Debugged Pro

Per month: roughly $8,333 to $16,667.
Per working day: roughly $400 to $800, if you count 250 working days.

For many small agencies in India, $8,000 a month is more than a senior developer's salary. For a US agency, it can match the cost of a full-time hire. That is the real comparison clients and owners will make.

Tan's use is personal and very heavy. He described agents running ahead of him and suggesting things, like "a Jiminy Cricket on your shoulder." Most teams will not need that level. Still, his bill shows how fast costs grow when agents run all the time with large context.

Why token bills climb so fast

From our own project work, costs usually rise for a few plain reasons:

Large context. Agents that read whole inboxes, drives or codebases send a lot of text with every step.
Many steps. One task can trigger dozens of model calls as the agent plans, checks and retries.
Parallel work. Running several agents at once multiplies the bill.
Premium models. The strongest models cost the most per token, and teams often default to them.

Limits are tightening too. We covered how Grok Bot usage limits were cut over AI agent costs. We also wrote about ChatGPT usage limits and the shared Codex token pool. Providers are clearly feeling the cost of heavy users as well.

How gstack Turns AI Agents Into a Virtual Team

Tan's gstack project is useful context for agencies. It is free, MIT licensed and public. Its GitHub page says it turns Claude Code and other coding agents into a "virtual engineering team."

The skills give the agent clear roles. These include a CEO who rethinks the product, an engineering manager, a designer, a code reviewer and a QA lead. There is also a security officer and a release engineer.

The lesson here is not "install gstack and relax." The lesson is structure. Agents do better with defined roles, clear steps and review points. That same structure also keeps spending under control, because each step has a purpose.

gstack has its critics too. When Tan first shared it in March, TechCrunch reported both strong praise and pushback online. So test it on a small internal project before you put it near client code.

Should Agencies Treat AI Agents Like Software or Like People?

Tan made one point that will split opinion. He said people should not treat AI agents like traditional software. In his view, some teams focus so hard on "repeatability and safety" that they lose what makes agents useful.

I understand the point, but agencies work under different rules. We handle client data, client code and client deadlines. A founder can take risks with their own notes. An agency cannot take those risks with someone else's systems.

So we take a middle path. Let agents be flexible inside a task, but keep the edges of that task firm. The OWASP guidance on excessive agency is a good reference here. It warns about agents with too much functionality, too many permissions or too much autonomy.

The data question nobody should skip

Tan's agent has read 20 years of personal notes. That is his choice to make. For a business, the same level of access to client email or shared drives would raise serious questions.

Before you connect an agent to any data, ask three things:

Does the task need this data? If not, leave it out.
Who else can see what the agent sees? Check the provider's data retention and training terms.
Can the agent act, or only read? Sending email, pushing code or paying bills should need human approval.

The NIST AI Risk Management Framework is a free, practical way to structure these checks. It works for small teams too, not just large companies.

A Practical AI Agents Budget Plan for Agencies

You do not need a six-figure token bill to get value from agents. Here is the plan we suggest to agencies and businesses in India and the USA.

Start with three to five tasks. Pick repeat work like code review, test writing, QA checks or research summaries. Measure the hours saved.
Set a monthly cap per project. Most platforms let you set spend limits or alerts. Use them from day one.
Match the model to the task. Use cheaper models for simple steps. Save premium models for hard reasoning or final review.
Trim the context. Feed the agent only the files and documents it needs. Smaller context usually means smaller bills.
Track cost per deliverable. Compare token spend against the time it saved. If a task costs more than a person would, stop it.
Review every month. Usage patterns change fast, and so do provider limits and prices.

If you want help building this kind of setup, our software development consulting team can review your workflows and suggest where agents make sense.

How to talk to clients about AI costs

Clients are starting to ask how agencies use AI. Be open about it. Explain which tasks use agents, where people still review the work and how costs are handled.

Some agencies fold token costs into their rates. Others bill them like hosting, with a monthly cap. Either approach works. What does not work is a surprise line item at the end of a project.

What This Means for Businesses Hiring Agencies

If you are a business owner, Tan's story is a useful reality check. Agent-driven work is real, and it can save time. But it is not free, and heavy use adds up quickly.

When you hire an agency, ask these questions:

Which AI agents do you use on my project, and for what tasks?
What data of mine will those agents access?
Who reviews the agent's output before it reaches production?
How are AI costs shown in my quote or invoice?

A good agency will answer these without hesitation. If the answers are vague, take that as a warning sign.

Key Takeaways

Garry Tan says he spends $100,000 to $200,000 a year on AI agent tokens, roughly $8,333 to $16,667 a month.
He admits the cost is too high for most people and says it must fall before personal agents go mainstream.
His agent has wide access to his personal data. Agencies should not copy that with client data.
Structure, like the roles in gstack, helps agents work better and keeps costs predictable.
Set spend caps, match models to tasks and agree on AI costs with clients up front.

Frequently Asked Questions

How much does Garry Tan spend on AI agents?

Tan told Walter Isaacson that he spends "maybe $100[,000] to $200,000 a year on tokens." TechCrunch reported the quote on November 2, 2026, ahead of the podcast episode going live.

What is gstack?

gstack is Tan's free, open source set of Claude Code skills, shared on GitHub under the MIT license. TechCrunch says it now has 23 skills. They give the agent roles such as product reviewer, engineer, QA lead and release manager.

Do agencies need to spend $200K a year to use AI agents well?

No. Tan himself called his spending a privilege that not everyone can afford. Most agencies can get real value from a few well-scoped agent tasks, a monthly budget cap per project and regular reviews of usage.

Is it safe to give an AI agent access to email and files?

It can be risky. OWASP lists excessive agency as a top risk for AI apps, meaning agents with more access or autonomy than they need. Give agents only the data and actions a task requires, and keep a human approval step for anything sensitive.

Should agencies charge clients for AI token costs?

Many agencies now treat token spend like hosting or software licences. Agree on it up front, set a monthly cap and report actual usage. That keeps the cost visible and avoids awkward surprises on the invoice.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He works with development teams and clients across India and the USA to plan, scope and deliver software projects. He writes about practical ways agencies can use AI tools without losing control of cost, quality or client data.