Garry Tan Spends Up to $200,000 a Year on Agents

Garry Tan spends up to $200,000 a year on agents

Garry Tan runs Y Combinator, the well-known startup program. On a new episode of Walter Isaacson's podcast The Builders, he put a price on his own agent use. TechCrunch reported that he spends about $100,000 to $200,000 a year on tokens. Tokens are the small units of text that AI work is measured and billed in.

His agents help with research and coding, and they also help run his day. The agent has read his email, his Dropbox, and his Apple Notes. It has even read his therapy notes going back 20 years. He calls it a Jiminy Cricket on his shoulder, one that knows him and runs ahead to flag what is coming.

His team is building the same idea at work. Y Combinator has an internal agent named Stray Light. It holds the context of every founder the program has backed and every application it has received. Tan has also shared his set of 23 Claude Code skills, called gstack, on GitHub for anyone to use.

Tan says the price has to come down before this fits most people. He called his own spending a crazy privilege that not everyone can do. He would rather see a personal AI that you control and program yourself, not one handed down by a big platform.

Try this on a smaller scale. Give an agent one job you can check, like sorting a folder or drafting a summary of notes you pick. Keep your inbox and your private notes out of it until you trust the setup.

Businesses paid Palantir more than the US government did

Palantir reported its third-quarter results on Monday, and one line marks a real change. The company's revenue rose 96% from a year earlier, to about $2.3 billion. Palantir built its name on work for governments and defense agencies, so the mix inside that total is the story.

US businesses paid Palantir $984 million in the quarter. US government work brought in $953 million. Business revenue passed government revenue for the first time in the company's history, and it grew 148% in a year.

Palantir raised its outlook for the full year to about $8.5 billion in revenue. Net income came to about $1.1 billion for the quarter. One quarter from one company does not settle where AI spending lands. It does show companies paying for AI systems at a scale that passed a defense-heavy business.

Anthropic made about $17 billion in one quarter

Anthropic filed an updated version of its paperwork to go public on Monday. The update adds the third quarter, and the growth is still steep. Revenue came to about $17 billion for the quarter, up 47% from the quarter before it. Across the first nine months of the year, revenue came to about $33 billion.

The costs are steep too. Anthropic swung to an operating loss of about $567 million for the quarter. The filing points to about $5 billion of short-term computing bought at higher prices than its long-term deals. Its purchase and lease commitments now add up to about $895 billion, stretching out over many years.

The filing also puts a number on how Anthropic builds now. In September, Claude wrote over 90% of the code contributions to Anthropic's own codebase. That is the company's own count, shared in a legal filing, not a sales demo. The tool the company sells is now doing most of the building at the company itself.

Then comes the warning. The filing says estimates for future models indicate the company may soon meet its own definition of dramatically accelerated progress from AI. That phrase covers AI speeding up the building of AI itself. No current model has crossed that line, and the company says it is still assessing the data. Anthropic says it does not expect to safely let Claude build later models fully on its own, without people supervising the work. It also admits it has missed a speed-up before and noticed only after it happened.

Also today, Anthropic added two sentences to Claude Code that tell it not to recreate a specific existing game, while original games still work, The Verge reported. OpenAI is merging the Chat and Work tabs in ChatGPT for a small group of users first, TechCrunch reported. MIT Technology Review covered a safety idea called reverse game theory, where the rules are set so that following them is the smart move for an agent. NVIDIA said cities are using its AI tools for planning, maintenance, and public services.

Today the money and the warning showed up in the same places. Agents are writing code, running workdays, and bringing in real revenue. The work ahead is keeping our checks as strong as the tools.

About this brief: This brief was researched and drafted by my personal AI assistant and approved by me before publishing. Product details change fast; check primary sources before you build on them.

Want more with the Tech Simplified crew? Subscribe on Substack.

Get the week's briefs in one email. The Briefcase packs Monday to Friday's AI Briefs into one Sunday email.

Follow me on LinkedIn for project updates, ideas for using AI, and lessons you can bring to your own work.