--- title: Personal AGI is an agent you own, not a chat you rent description: Garry Tan says AGI won't arrive as one big announcement. It's already here, spread out, as an agent that runs on your own machine and works from your own notes. What he means by personal AGI, how he builds his, and why he says you should own your skill files. date: 2026-11-09 tags: AI agents, personal AGI, agent skills, agent memory, Y Combinator --- "AGI isn't arriving as an event," says Garry Tan, who runs Y Combinator (YC). It arrives spread out, as your own agent, working from your own context and doing your work. He calls it personal AGI: general intelligence for one person, not for everyone at once. Most people wait for AGI as a headline: one lab, one announcement, one day. Garry's point moves the question from the labs to each of us. If the gain comes from what an agent knows about your work, you can start building it tonight. And nobody can ship it for you. He made the case at [Startup School 2026](https://www.youtube.com/watch?v=eRrc1pUY5oU), in a 42-minute talk called *Own Your Intelligence*. He opens with Spinoza, the seventeenth-century philosopher who said God isn't a king on a throne but is spread through everything. Garry's version for today: people watch the sky for a god in a data center. What they wait for is already in the room. It looks like a terminal window, a folder of Markdown files and a job that finishes while you sleep. ## A subscription you rent is not yours Garry is careful with the words, because "personal AI" is already a marketing label. He doesn't mean a chatbot at $20 a month. He doesn't mean a better autocomplete, or an assistant that knows your calendar and nothing else. Those, he says, are rented. They reset when you close the tab. They know what everyone else knows. And when the company behind them changes course, your assistant changes on someone else's schedule. Personal AGI is the other kind. It runs on your own infrastructure, reads a memory you own and follows procedures you wrote. It gets better every day you use it, because every day it knows more of your life. A rented assistant only improves when the company ships a new version. In his words, one is a product you consume, the other is an asset you build.

Two kinds of personal AI

A rented assistant, or an agent you own

RentedOwned

Where it runsOn the company's serversOn your own infrastructure

What it knowsWhat everyone else knowsYour life and work, from a memory you own

What it followsThe company's productProcedures you wrote

When you close itIt resetsIt remembers

When it gets betterWhen the company shipsEvery day you use it

In his words: a product you consume, or an asset you build.

From the talk. The row names are mine.
## Same model, 2x or 100x His formula has three parts. First, a frontier model: you rent it, everyone can buy it, and it gets cheaper every quarter. Second, your context: what you know about your work, which nobody else has. Third, a harness, the program around the model that gives it tools and runs it step by step. He names OpenClaw, Hermes Agent, Claude Code and Codex. The model is the same for everyone. That's how he explains a gap he sees: with the same Claude and the same API, some people get two times more done and others a hundred times. The difference isn't in the model. It's in the context you give it, how relevant it is, and whether it arrives at the right step. He gives one sign from YC's own companies. In the Winter 2025 batch, a quarter of the startups had code that was 95% written by AI. That batch is on track to be one of the fastest-growing YC has funded. He says himself that he can't prove the AI caused the growth. What he sees is that the fastest founders treat AI as a workforce, not as autocomplete.

His formula for personal AGI

The model is rented, the context is yours

RentedA frontier modelEveryone can buy it, and it gets cheaper every quarter.
OwnedYour contextWhat you know about your work. Nobody else has it.
The wiringA harnessOpenClaw, Hermes Agent, Claude Code or Codex.

Same Claude, same API: some people get two times more done, others a hundred.

His three parts, from the talk. The two times and the hundred times are his words, not a measure.
## Seven things in your head, a library behind the agent A person holds about seven things in mind at once, give or take two. It's the most famous result in cognitive psychology, and the reason phone numbers used to have seven digits. Garry's point: every checklist, org chart and stand-up meeting is a crutch for that limit. An agent holds a million tokens, about a thousand pages. That's three Harry Potter books open at once, searchable in seconds. But a thousand pages is also very little. Your life isn't three books, it's a library: every email, meeting, decision and conversation. So the real question is who picks the three books open on the desk. His answer is a library plus a librarian. His own library holds about 220,000 Markdown pages, 25 years of his life: emails, meetings, notes and drafts. Agents write it, sort it and search it. When a founder emails him about a crisis, his agent has pulled every past conversation with that founder before he finishes reading. His strongest example isn't his own. A friend of his has a son with a rare form of epilepsy. With no lab and no grant, the father built 80,000 Markdown files on his son's condition. Every specialist visit, paper, seizure log and drug interaction is in there, linked together. When a new doctor has an idea, he knows in minutes whether it was already tried. "That is personal AGI," Garry says.

Working memory, then a library

A thousand pages on the desk, a library behind it

7things a person holds in mind at once, give or take two

1,000pages an agent holds, about a million tokens

220,000Markdown pages in Garry's library, 25 years of his life

The real question: who picks the three books open on the desk.

His numbers, from the talk. The thousand pages is his conversion of a million tokens.
## A skill is a page of English The procedures his agent follows are skill files: Markdown pages that say how to do one job. He reads one out, lightly edited. When a meeting recording arrives, transcribe it with the speakers' names. Pull out each commitment, who made it and the deadline. Link every person named to their page in the library. File the summary and the transcript. If something contradicts what we already believe, flag it, don't overwrite it. That's the whole skill. His test is simple: if a smart intern could follow it, an agent can run it. So writing clear English is programming, and the model is the compiler. At YC, he says, people in media, events and finance who never opened a terminal now write skill files and scheduled jobs. One person in finance turned about a hundred Excel workbooks into a single app. He adds a rule that, he says, explains every agent failure he has seen: know where the computing happens. Judgment, taste and reading what someone really wants belong in the model, steered by a Markdown file. Arithmetic, database queries and schedules belong in code. An agent can seat five people around a table in its head. To build schedules for 6,000 people at a conference, it has to write code.

Where the computing happens

Judgment in the model, arithmetic in code

In the modelIn code

What it doesJudgment, taste, what someone really wantsArithmetic, queries, schedules

How you steer itA skill written in MarkdownScripts and a database

His exampleSeat five people around a tableSchedules for 6,000 people at a conference

His split. In the talk, the two sides are latent space and deterministic space.
## Start with one folder His library didn't start at 220,000 pages. It started as a folder with a few Markdown files about the companies and people he dealt with. "Nobody builds the warehouse first," he says. Then he gives five steps. - **Pick a harness.** Run an agent on your own machine. Any of the four he names does most of this, and his own, GBrain, is free and open source. - **Start your library.** One folder of Markdown files: a page per project and per person you work with, with what only you know. - **Write your first skill.** Take the weekly task you hate most. Explain it the way you would to a smart friend on their first day, then correct it each time it gets it wrong. - **Make it a recurring job.** Every morning at seven, do this. Every Friday, sum up that. - **Never do one-off work.** At the end of each task, ask the agent to turn what it did into a skill. His rule at YC: if you have to ask for something twice, you failed. He's honest about the curve. Week one is a toy: the library is thin and you fix more than you save. By week four, the agent starts answering with your context. By week twelve, a dozen skills run the parts of your week you used to dread. Most people, he says, quit in week two. He also gives a warning. A library nobody looks after is "a garbage dump with great search". It serves an old fact with full confidence, and a bad skill repeats a bad process forever. So every fact needs a source, new facts get checked against old ones, and someone, or some agent, has to prune.

His five steps

Start with one folder, and never do one-off work

  1. Pick a harnessRun an agent on your own machine.
  2. Start a libraryOne folder of Markdown: a page per project and per person.
  3. Write a skillThe weekly task you hate most, in plain English.
  4. Make it a jobEvery morning at seven, or every Friday.
  5. Never do it onceTurn each finished task into a skill.

Week one, a toy. Week four, it answers with your context. Week twelve, a dozen skills. Most people quit in week two.

His steps, shortened. The weeks are his estimate, not a measure.
## Own your skills, or your job becomes one Then the talk turns political. A skill file, he says, isn't a document. It's a piece of how you think, taken out of your head, written down, and ready to run. The same file leads to two opposite futures, depending on who controls it. His example is made up: Maya, a support engineer. Over two years, she teaches her agents 40 skills. How to handle a critical incident at two in the morning. How to calm a customer who is about to leave. How to write a postmortem that prevents the next incident. In the first version, the files live in her own repository. She changes jobs and starts the new one with two years of judgment ready to run. In the second version, they live in the company's repository. She leaves with nothing, and the company keeps running her judgment without her. His rule: "Own your skills because if you don't, your job becomes a skill file." He compares it to craftsmen, who owned their tools until the factory came, where the loom belonged to the mill. Knowledge workers felt safe because their tools lived in their heads. Skill files change that. His answer is written for founders: start a company, and keep your skills in a repository you control from day one. For an employee, it's harder. Most employment contracts say that what you write for your job belongs to the employer. So read yours before you copy a skill file out of a company repository.

Maya's 40 skill files, his made-up example

The same files, two different careers

In the company's repoIn her own repo

When she leavesShe leaves with nothingThe files go with her

Her judgmentKeeps running without herReady on day one of her next job

Each year she worksAdds to the company's filesAdds to her own

His rule: own your skills, or your job becomes a skill file.

Maya is the talk's fictional support engineer. The two columns are his two versions.
## Three objections he answers He ends with the three objections he hears most. **Better models will make this useless.** He says the opposite happens. The better the models get, the more the difference moves to context. A smarter reader gets more out of the same books, so each new model makes your library worth more. **It's just RAG.** RAG, short for retrieval-augmented generation, means fetching documents and handing them to the model. "Sure. And Postgres is just B-trees," he answers. Fetching is the easy part. The hard part is what gets written down, how it's linked, what stays close at hand, and who decides when two facts disagree. **It will leak.** Putting your whole life in one system sounds risky. His answer: your life is already spread across ten clouds, owned by companies whose interests aren't yours. Keeping it on your own machine, in your own repository, under your own keys, means taking custody of it. It doesn't create the risk.

Three objections, as he answers them

Each answer comes back to context you own

Better models will make it uselessContext matters moreA smarter model gets more out of the same library.
It's just RAGFetching is the easy partWhat gets written, linked and trusted is the hard part.
It will leakCustody, not new riskYour life is already spread across ten clouds.
His answers, shortened.
## Try it The definition of personal AGI starts at about 4:46 in the video, and the five steps at 24:16. - Y Combinator, [Garry Tan: Own Your Intelligence](https://www.youtube.com/watch?v=eRrc1pUY5oU), Startup School 2026, and [the transcript](https://www.ycrootaccess.com/p/garry-tan-own-your-intelligence). - [GBrain](https://github.com/garrytan/gbrain), his library and skills, free and open source. - [Spending more tokens only pays with good skills and real tests](/notes/tokenmaxxing-fat-skills-real-tests.html), on an earlier talk where he explains "thin harness, fat skills". - [Intelligence on tap](/notes/intelligence-on-tap-consumer-ai.html), on what to build while AI is still expensive. - [Fork your agent's harness, keep the winner](/notes/fork-your-agents-harness.html), on improving the harness itself. Then try his second step this week. Make one folder and write one page about one project. Put in what you're building, who cares about it, what you owe and what was said last time. Ask an agent a question that only that page can answer. If you try it, tell me what it got right.