Garry Tan on Personal AGI, Skill Files and Owning Your Context

Key points

1. 1

   The model is rented, your context is owned

   Tan's formula for the next decade: a rented frontier model that gets cheaper every quarter, plus context nobody else has, plus a harness that wires them together. He says a 2x worker and a 100x worker use the same model. The gap is what context the agent gets, and whether it arrives at the right step. A $20-a-month chatbot resets when you close the tab. An agent on your own memory improves every day you use it.

   “Model quality is rented, but your brain is owned, ideally by you.”

2. 2

   400x his 2013 output, 8x at the floor

   In 2013, as a YC partner building Bookface at night, Tan shipped about 14 useful lines of code a day. Running YC full-time now, he puts his output at about 400x that. Then he deflates it. Assume bloated code, half of it scaffolding, and a flattering count, and it is still 8x at the floor. He says the same multiple applies to design, product and growth work.

   “I did the math on my output, and I'm at about 400x what I did in 2013.”

3. 3

   A skill file is a page of English

   His setup is mostly skill files: plain-English instructions an agent follows, which he sums up as fat skills, thin harness. His real example handles meeting recordings from CircleBack. It transcribes with speaker labels, pulls out commitments and deadlines, links each named person to their page and flags contradictions. The test is whether a smart intern could follow it. At YC, finance and events staff now write them. One finance person turned about a hundred Excel workbooks into a single app.

   “And that's the test actually. If a smart intern could follow it, an agent can run it.”

4. 4

   Split latent work from deterministic work

   Tan says the first question for any agent is where the computing happens. Taste, judgment and reading a vague request live in the model, steered by a text file. Arithmetic, database queries and schedules belong in code the file calls. Seating five people is a model job. Custom schedules for 6,000 event guests needed the agent to write code.

   “where is the computation happening? And there are exactly two answers, and confusing them causes every agent failure I've ever seen.”

5. 5

   Own your skill files, or your job becomes one

   A skill file is a piece of your judgment, written down and runnable, so who holds it decides who benefits. Tan's made-up support engineer, Maya, teaches her agents 40 skills over two years. If the files sit in her repository, she starts her next job with years of judgment on tap. If they sit in the company's, the company keeps running her judgment after she leaves. His advice: keep them where you control them from day one.

   “I believe skill files are yours. Own your skills because if you don't, your job becomes a skill file.”

6. 6

   Revenue per person that did not exist before

   Tan's proof that company math changed is Emergent, from YC's Summer 2024 batch. It went from launch to nine figures of revenue in eight months, and had 15 people when it crossed $15 million a year. Another Winter 2024 company reached $60 million a year with about 40 people. He says a quarter of the Winter 2025 batch had code 95% written by AI. He admits he cannot prove AI caused the growth.

   “When they crossed $15 million in annualized revenue, they were 15 people.”