1 · Sensor layer
Channels that bring signal in from the real world.
A company that improves while you sleep
Extract your company's know-how, make it legible to AI, and wire up loops that get better on their own — even overnight.
“If it was recorded, it happened — to the AI. If it was not recorded, it did not happen.”
5×Revenue / employee vs 18 months ago
2000+Hours of office hours recorded
150Pages in the regenerated manual
5Layers in the AI loop
Companies are still built like Roman legions
And AI breaks the assumption that hierarchy is how value gets made.
The Roman legion was designed to project power across two continents — from Rome all the way to Hadrian's Wall in Scotland — through nested hierarchies with fixed spans of control, where named individuals passed orders down and sent information back up. Most companies today are organised exactly the same way: human beings are the conduit for information flowing up and down.
Copilots are the wrong mental model
A year ago, people described AI as productivity: copilots that make an engineer 20% faster, tools bolted onto existing workflows to ship a bit more software. But that just straps a more powerful engine onto the old way of working. The bigger move is to reimagine what a company even is.
Extract the domain knowledge
Every company has know-how scattered across people's heads, Slack threads, emails and Notion docs — together they define how the company works. Make that knowledge legible to AI and you can shift from a hierarchical org to an intelligent, AI-native one: a set of recursive, self-improving AI loops that keep improving even while you sleep.
The five layers of an AI loop
Run every step with minimal human intervention and the system gets better on its own. Tap a layer for detail.
2 · Policy layer
Rules for what AI may do, what needs a human, what must be logged.
3 · Tool layer
Deterministic APIs the AI can call — query a DB, read a calendar.
4 · Quality gate
Evals, safety filters, human review for high-risk actions.
5 · Learning mechanism
Catch what didn't work in the real world and loop back to the top.
The “holy shit” moment at YC
A live, in-production example of a loop that learned to fix itself overnight.
- Step 1
A deterministic query agent
It started simple: an agent with deterministic tools to query our database — “when did I last have office hours with this company?”
- Step 2
It got smarter — but still a sidekick
Using RAG and richer queries, it could surface five relevant founders to introduce — say, anyone in petrochemicals. Useful, but this was last year's model: a sidekick making me 20–30% more effective.
- Step 3
We put a monitoring agent on top
It watched every query every YC employee ran — which worked, which failed — and asked why the failures failed. Different tools? An updated skills file? A new database view or index?
- Step 4
It shipped the fix overnight
Overnight it wrote the code, opened a merge request to the YC codebase, had another agent review it, merged, and deployed.
- Step 5
Next morning, the query just worked
A human asked the same question the next day and it succeeded. That was the holy-shit moment — not AI making you 20% faster, but AI running the whole loop to figure out how to improve itself.
The same logic, everywhere
Find the parts of your company that can run as loops, put humans in a supervisory seat, and just keep feeding it tokens.
A self-optimising product loop
Find funnel friction, A/B test it, ship the winner — repeat.
Customer-suggestion triage
An agent acting as CPO + CTO decides what to build and ships it overnight.
Every function as a loop
Reframe each part of the company as a recursive, self-improving loop.
Two shifts this forces
What changes once you take loops seriously.
Burn tokens, not headcount
Companies are reaching Demo Day with ~5× the revenue per employee they had 18 months ago, and it will keep going through Series A and B. Soon you'll be constrained by token usage, not headcount. The blunt proxy today is everyone's token usage — gameable the moment it becomes a leaderboard for promotions, but directionally it tells you who is “token-maxing” and who isn't.
Middle management is over
Middle management existed to solve a coordination problem — and AI can do that now. Everyone becomes an IC (a builder or an operator), and crucially, every effort has a single Directly Responsible Individual: a named human, not a committee.
Three roles, no permanent managers
Jack Dorsey's framing: the world model handles alignment, so people sit closer to the work and the customer.
Individual Contributor (IC)
Deep specialists who build and operate a specific layer.
Directly Responsible Individual (DRI)
Owns a cross-layer problem and the customer outcome.
Player-coach
Builds and grows people at once — replaces the info-routing manager.
Where to start: make the org legible to AI
Record everything, distil it, and treat software as disposable — but data as sacred.
First, record everything. Every partner email now lands in the YC database; every Slack message, every DM, and every office hour from the last few months is recorded. The principle is simple: if it was recorded, it happened — to the AI. If it wasn't, it didn't.
Distil it down (diarisation)
You can't pump 100,000 hours of recordings into a context window, so you distil: aggregate, synthesise into the important parts, and leave the AI breadcrumbs back to the raw source.
Regenerating the user manual
YC's user manual was written 5–10 years ago and is partly out of date. Over one weekend, Haj used ~2,000 hours of recent office hours to regenerate it: a 150-page manual, dramatically better, now updatable every month. Every new piece of advice is compared with the manual and either folded in or discarded — a living brain of what 16 partners actually tell founders.
Software is ephemeral, data is sacred
- If something produces a self-improving artifact, keep it; if not, throw it away.
- Every function can now one-shot its own dashboards and internal tools — treat that software as disposable.
- Store the data preciously (Garry keeps every email as markdown); regenerate the software when the models get smarter.
So what are humans for?
Humans live around the edge of the company brain, where intelligence meets reality.
All your data, emails, DMs, skills and know-how — that is the company brain. Humans sit around the edge, interfacing with the real world.
Humans reach into the places the models can't go yet: novel situations, ethical calls, high-stakes moments — like a founder thinking about breaking up with their co-founder.
Sales conversations need a real person in the room for the next twenty years.
Chapters
How the talk unfolds, start to finish.
- 00:00
Companies Are Roman Legions
Why most orgs still route information up and down a hierarchy.
- 00:54
Copilots Are the Wrong Mental Model
Productivity gains are just a stronger engine on the old way of working.
- 01:55
Extract the Domain Knowledge
Make the know-how in heads, Slack and email legible to AI.
- 02:24
The Recursive Self-Improving Loop
Sensor → policy → tool → quality gate → learning.
- 04:12
The Holy Shit Moment at YC
A monitoring agent that fixes failed queries overnight.
- 05:50
Self-Optimizing Product and Support Loops
A/B tests and customer triage that run themselves.
- 06:29
Burn Tokens, Not Headcount
Revenue per employee is up ~5×; the constraint is tokens.
- 07:23
Middle Management Is Over
Everyone is an IC; every effort has a DRI.
- 08:05
Make Everything Legible to AI
Record everything — emails, Slack, DMs, office hours.
- 09:40
Regenerating the YC User Manual
2,000 hours of office hours → a 150-page living manual.
- 11:19
Software Is Ephemeral, Context Is Valuable
Cherish the data; regenerate the software as models improve.
- 12:18
Where Humans Still Matter
Novel, ethical, high-stakes moments — and sales.
Sources & corroboration
The original talk and the references behind it.
This page is an unofficial, fan-made summary for study. For anything load-bearing, go to the original sources below.
VideoHow to Build a Self-Improving Company with AI — the talkY Combinator · YouTubePost“Recursively self-improving companies” — Tom's own breakdownTom Blomfield · XReferenceThe Playbook for Building an AI-Native CompanyDiana Hu · YC Startup LibraryReferenceFrom Hierarchy to Intelligence — the IC / DRI / player-coach rolesJack Dorsey · BlockEssayYour Company Is an Intelligence — You Just Haven't Built It That Way YetPete SenaProfileTom Blomfield — writing & talkstomblomfield.com
Would you build it this way today?
Most companies are still small enough to build in this shape from day one — a recursive, self-improving intelligence with humans around the edge. There's no excuse not to.
Watch the full talk