An AI-native agency installs a nervous system, not a deliverable
by Ygor Fonseca, Founder & Systems Lead
Garry Tan, the CEO of Y Combinator, recently published an essay describing his personal AI setup. He keeps a structured knowledge base of roughly 100,000 pages in a git repo, more than a hundred reusable skill files, around a hundred cron jobs running, and a meta-skill called Skillify whose job is to turn one-off workflows into permanent ones. He fact-checks outputs across multiple models. He keeps a page on every person he meets. After every meeting, the system walks through every person and company mentioned and updates their pages with what was discussed. He calls the difference between this setup and a normal AI workflow the difference "between having a filing cabinet and having a nervous system."
This post is about what that essay means if you are running a brand instead of running yourself. Garry's architecture — a thin harness, fat skill files, fat data, and a meta-skill that turns repetition into infrastructure — is the right shape for any operator who wants AI to compound rather than just answer questions. The translation from one-person setup to a brand operating with twenty or two hundred people is mostly mechanical. The harder part, for most brands, is installing it. That is the work an AI-native agency exists to do.
The smuggled assumption inside most "AI agency" engagements is that the agency runs the AI on the brand's behalf, the brand sees outputs, and the system itself is the agency's private infrastructure. When the engagement ends, the system ends. That is a campaign, not an install. The right shape leaves a working system behind, in the brand's repo, on the brand's data, with the brand's keys.
The argument is older than the essay
Karpathy, in his Sequoia talk on Software 3.0, framed software as something humans now build by orchestrating models, sensors, and actuators rather than writing every instruction by hand. Anthropic's Building Effective Agents paper laid out the workflow patterns underneath — chains, routing, parallelization, orchestrators, evaluator loops. Emergence Capital and Sequoia, on the investor side, have been arguing for over a year that services delivered as software is the dominant business model of the next cycle. None of these are saying "use AI more." They are all saying the same thing from different angles: the value is in the assembly, not the raw model.
Garry's contribution is showing what the assembly looks like at the scale of one person. The architecture is plain. A thin harness that routes. Fat skill files that each do one job well. Fat data that grows daily. A meta-skill that turns repetition into permanent infrastructure. Run that for a year and the system you have is not replaceable by a chat window. It is replaceable only by another system of the same shape.
The same architecture, scaled sideways
A brand operator has the same raw materials Garry has, only more of them. Every campaign brief, every customer email, every retention flow result, every paid-media debrief, every QBR doc, every performance dashboard — those are all pages in a brain that does not exist yet. Most brands keep that knowledge scattered across notes, docs, and senior people's heads. When someone leaves, a chunk of it leaves with them. When a new agency starts, they spend two months learning what the last agency had already learned and forgot to write down.
A working AI-native setup at brand scale is the same shape Garry described, pointed at the brand. A repo for the brand. Pages for every customer segment, every retention cohort, every paid channel, every supplier, every product launch. Skills that handle the workflows the brand actually runs — the weekly retention review, the launch-prep brief, the new-creative QA pass, the post-promo debrief. A meta-skill that turns the next manual workflow into a permanent one. That is the asset.
Why brands rarely install this themselves
Garry built his system because he is technical, has time after midnight, and treats the build as a personal project. Most brand operators have none of those three. They have an ecommerce store to run, a paid-media plan to hit, a launch calendar that does not move, and a finance team waiting on the numbers. They do not have a head of growth who also writes skill files at 2am. The architecture is right. The build cost, in time and skill, is the bottleneck.
This is the gap an AI-native agency closes. Not by running the AI for the brand, but by installing the architecture inside the brand's own repo and teaching an internal owner to extend it. The agency builds the harness. The agency writes the first set of skills against the workflows the brand already runs. The agency seeds the brain by ingesting the brand's existing docs, dashboards, and meeting notes. The agency sets up the meta-skill that converts the next manual workflow into a permanent one. And then the agency steps back to the role of operator and reviewer, while the system runs inside the brand's stack.
This is what we call Codified Engagement. The brand ends up with a working system, in their repo, on their data, with their keys. The agency leaves a compounding asset behind, not a stack of decks.
What good agency support looks like in this model
Three questions worth asking any agency you are considering for AI-native work, framed as the answers a strong partner gives.
Where does the system live, and who has root? A strong AI-native agency installs into the brand's own repository, on infrastructure the brand controls, with named internal owners holding root access from day one. The repo is shared from the start, not handed over at the end. If the answer instead is "our platform" or "our SaaS," the engagement is rental rather than installation, and the architecture stops compounding the day the contract closes.
Is there a meta-skill, and will you teach an internal owner to run it? The single move that turns a project into a system is the meta-skill that takes today's manual work and converts it into tomorrow's permanent skill. A strong partner builds that meta-skill into the engagement from week one and runs a knowledge transfer to an internal owner — usually a senior operator on the brand side — so the loop continues running after the agency hours end.
At month six, what is running without you? A strong partner has a clean answer here. A list of skills, crons, and brain pages that keep producing work after the contract closes. The honest version of the question is whether the engagement was a campaign or a system. The answer should be visible in the brand's repo, not described in a slide.
Where this leaves us
Garry's essay is the clearest public description of what compounding AI looks like at the scale of one person. The architecture he describes — thin harness, fat skills, fat data, meta-skill that compounds — works at brand scale too. Most brands cannot install it alone, which is exactly the work an AI-native agency should be doing: installing the system inside the brand's stack, training an internal owner to extend it, and stepping back into the operator-and-reviewer role while it compounds.
The right test of an AI-native engagement is not what gets shipped while the contract is active. It is what is still running in the brand's repo six months after the agency stops billing.
Talk to us about installing a brand nervous system.
Citations: Garry Tan, "Meta-Meta-Prompting: The Secret to Making AI Agents Work" (the personal AI setup and the filing-cabinet-versus-nervous-system framing); Andrej Karpathy, Sequoia Ascent talk on Software 3.0; Anthropic, Building Effective Agents; Emergence Capital, AI-Native Services Playbook; Sequoia / Konstantine Buhler and Sonya Huang, Services: The New Software. Companion pieces: the named-method post on why the artifact has to be on the public site, the gap is not your model on what compounding looks like in our own delivery, and the AI-native marketing team 3-role chart for what the brand-side operating model looks like.