Orionfold Flow is coming to Mac · patent-pending AI agency for documents
Join the waitlist →
One email a week until launch, no more.
Conduct beautiful documents with AI agency built in.
Your text changes two ways only: an edit you typed, or a change you approved.
I used to think becoming AI-native was a membership.
A batch. A software team that had been pushing models for two years. A warehouse of recorded founder calls. An orange logo you could put on the about page.
Then I watched a YC partner spend twenty-five minutes explaining how Y Combinator is turning itself into an AI-native company.
Not a portfolio company. YC.
The talk is on YouTube as Building And Structuring An AI Native Company. He opened with the only honest sentence in this whole genre:
No one knows how to do this. If anyone tells you they have it figured out, they are probably lying.
Then he described the work anyway. English to SQL over 7,000 companies and 20,000 founders. A second agent that runs overnight, reads every failed query from the day, and files pull requests so the same question works tomorrow. Three or four thousand hours of recorded office hours, mined into a 500-page user manual that used to be rewritten by a procrastinating partner and is now rewritten by the advice the partners actually give. An agent that can answer a Slack question in Tom's voice, or Niccolo's, or Harj's, because it can recall sixteen partners at once.
He called the first time he saw it a head explosion moment.
I know that feeling. I have been living a smaller version of it at one desk in Sacramento, building Orionfold, writing the receipts down, and putting the same loop in the document.
This is the sequel I promised myself after Limitless, Without the Pill. That piece was about one person becoming more orchestrated. This one is about what happens when the company itself starts to learn.
And why that future is not reserved for people who already have the badge.
The Roman tent is still the org chart
The partner started in an unexpected place. The Roman legion.
The smallest unit was the contubernium: eight soldiers, one tent, one mule, one decanus. Ten of those made a century, which was eighty men, not a hundred. Information went down. Reports came up. A human being was the conduit the whole way. That structure projected power from North Africa to Hadrian's Wall.
Two thousand years later, he said, we are still running the same kind of thing. The labels changed. Director. VP. Slack channel. Weekly. The job did not. A person is still the pipe.
Jack Dorsey kicked the thought into public about two months before the talk. The underlying assumption is that organizations have to be hierarchically organized with humans as the coordinating mechanism. AI breaks that assumption. You do not have to use people as the router anymore.
Most companies have not noticed. They met ChatGPT as a Q&A bot. Question in, answer out. Then they met agents, which go away, get stuck at 3 a.m., and wait for a human to unstick them. The human is still the gate. The company is still the legion. The tent just has a chatbot in it.
You can make an engineer 20 percent faster this way. You can give a lawyer a copilot. You can ship more software. You are still paying people to carry information up and down the rungs.
The partner's question was different. What if the company is a series of self-improving AI loops, built from the ground up?
What a loop actually is
He drew it simply enough that I wrote it down.
At the top: the real world. Product telemetry. Support tickets. Billing signals. Code changes. Inbound mail.
Then a policy layer. What is the AI allowed to do. What must it ask approval for. What must it log.
Then tools. Internal APIs. Mail. Billing. MCP. Whatever the company already has.
Then quality gates. A human, if the cost of being wrong is existential. More often a second, adversarial model. Is this prompt injection. Is this financial advice we are not allowed to give. For engineers, a second model doing code review.
Then learning. Deploy. Watch the world. Keep the change that went up the hill. Discard the one that went down.
If that whole loop can run without a person sitting in the middle, the product improves while you sleep.
YC already has a few of these. PostHog, in a talk the same day, had the same shape for product surfaces: telemetry in, pull requests out, eventually auto-merge, against a vision document that says what is in scope. A researcher named Copacetic had applied it to machine learning. Come up with an idea. Test it overnight. Hill climb. Computers are good at hill climbing because they do not have to sleep.
The office-hours story is the one that stayed with me.
The internal user manual is 500 pages, written over fifteen years. A lot of it was excellent five years ago. Then AI happened, and a pile of the advice stopped being true. The partner was supposed to rewrite his section. He procrastinated. They started recording office hours. Six months later they had thousands of hours. Harj's idea: transcribe it, mine the advice they actually give, and let that rewrite the manual.
When the advice changes, the manual changes. Then you make the manual queryable. A founder asks, in Slack, how to price the first customer, or what to do when the third one is about to churn. YC has answered versions of that question 500 times. The agent can now give you Tom, or Niccolo, or Harj, with perfect recall.
That is a company becoming legible to itself.
Then you give the agent a virtual machine. Search the web. Crawl the directory. Read Slack. Write a plan to disk so a failure halfway through can resume. Write code. Run the code. Compare it to the plan. Repeat. He called that an AI employee. OpenClaw and Hermes are the early, slightly chaotic versions: a VM with an agent living inside on a repeating loop.
The loops start talking to each other. People in the audience were already wearing t-shirts that said company brain.
His definition, stealing Dorsey's words: in a conventional company, intelligence is spread through the people and the hierarchy routes it. In the new model, intelligence lives in the system. People live at the edge, where that intelligence touches the real world. They sense the feeling in a room. They make the ethical call. They visit the client office. They pitch the investor. They do the work a model should not do alone.
They stop being the pipe.
This is where the story usually becomes a club
If you stop the video there, it sounds like a membership benefit.
You need 7,000 companies of training data. You need 16 partners. You need a software team that has been pushing LLMs for two years. You need every founder call recorded. You need permission to ban Slack DMs so the AI can read the work. You need to be early-stage enough, or YC enough, to "just build it right from the start."
The partner said that last part out loud. Most of the people in the room were early enough to do it. The rest of the economy is not in that room.
The PE operating partner in Ohio is not going to stand up a VM farm this quarter. The comms lead at a 200-person company is not going to ban DMs. The founder raising from a kitchen table does not have 4,000 hours of office hours. The lawyer, the consultant, the analyst, the chief of staff: they have documents. They have a name on those documents. They have a Tuesday.
If AI-native work only works for YC, it is a demo.
I do not accept that.
The partner's own practical advice, the part he is already giving founders, is smaller than the sci-fi:
Burn tokens, not headcount.
Everyone is an individual contributor. Come to the meeting with a working prototype, not a deck.
Give one person a DRI. A single head on the block. Committees grind things to a halt. He learned that at a bank.
Make everything legible. Record the meeting. Transcribe it. If the AI cannot see the Slack DM, ban the DM. Every action has to leave an artifact. If it did not leave a written or recorded version, it did not happen as far as the AI is concerned.
That last sentence is the whole product I am building.
The missing object is the file
YC can make the company queryable because they already treated the work as data. Applications. Office hours. Investor calls. The user manual. Pull requests. Success and failure on yesterday's queries.
Most knowledge work does not look like that.
It looks like a proposal in a chat canvas. A summary that got pasted into email. A deck that no longer matches the model. A Notion page that drifted. A "final_v7_really_final" in someone's Downloads folder. The interesting decision happened in a meeting and died there.
Chat is where work evaporates. I have written that on the Flow page because I keep watching it happen.
The partner said every action needs to create an artifact. I agree. I am more specific about which artifact.
The document is the workplace.
Not the chat. Not the canvas. Not the export that breaks the formatting and loses who approved what. The file on disk. Plain Markdown, in a folder you already have. No import. No library format. Every other tool on the Mac can read the same bytes.
If the company brain cannot read the work, it is not a brain. It is a pile of sessions.
Flow is a native Mac app where AI does the work in that file. The change shows up as a diff you approve. A receipt stays in the file. Nothing is saved until you approve it.
That is the policy layer, the quality gate, and the learning record, sitting where the name on the document already lives.
A silent rewrite is not a feature
The partner was careful about gates. A human should still make the call when the cost of being wrong is existential. Ethics. Novel situations. High-stakes moments.
Knowledge work is full of those moments, and they do not announce themselves as "existential." They look like a code comment. A citation. A number in a pitch. A quote in a press note.
While I was building Flow I tested the Mac's own Writing Tools on a document that held a code comment with two typos. The system feature corrected and capitalized the text inside the code block. Silently. No diff. No approval. No record. On a real working note, that is corruption of code.
I measured it. I dated it, 2026-07-30. I reproduced it. Then I retired Writing Tools inside Flow and replaced the shape.
Five actions. Proofread. Summarize. Translate. Convert Text to Table. Convert Table to Text. Every one walks the same path.
It says what it will do, read from the proposed bytes, not from the model's word.
It shows the evidence, or admits there is none. A summary is deliberately not scored. Flow prints that instead of inventing a number.
It shows the exact bytes. The diff. Line by line. A run that started before your edit cannot land after it.
You approve, or you decline. Decline changes nothing at all.
A silent rewrite is not a feature. It is a risk.
YC's quality gate, for a bank, is "are we giving financial advice, yes or no." Flow's quality gate, for a person whose name is on the file, is the same idea at the scale of a sentence. Named checks. Six honest states. An override never turns Failed into Passed. The failure stays visible.
That is how you let the loop run without turning the human into a spectator.
One knowledge worker, one loop
You do not need sixteen partners to start.
You need one recurring document and a refusal to let the work die in chat.
A founder raising: the pitch is a living file. Every claim can keep its source. When a partner asks where a number came from, the answer is attached to the number. One fake citation can kill a round. The partner in the YC talk described pumping forty or fifty investor calls into an AI to debug the pitch. You may not have YC's side of those calls. You can still keep the pitch itself honest.
A comms lead: journalists now screen for slop. A story with quotes, stats, and coverage attached earns a kind of trust a polished paste job cannot. A human approved every line. The receipt says so.
An independent operator: the proposal, the analysis, the brief. Sensitive pages stay on the Mac by default. The record shows what ran and what it cost. Local is free because it is free. A billed run I actually paid for recorded $0.00425, 105 input tokens and 149 output, exact decimal arithmetic. That is not a typical figure. It is one receipt.
An enterprise AI owner: the failure mode is not "we lack models." It is "we cannot tell which 30 percent is wrong, who approved it, and what it cost." Flow's answer is structural. Put the evidence, the approval, and the receipt at the source of the work, inside the document.
Bring open AI models to those documents. Stay in control of cost, privacy, and quality on your Mac. Switching to a frontier model is a checkbox. Flow selects the best model for the job. Four domains decide where the work may run: Local, LAN, Cloud prepaid, Cloud postpaid. Each has its own switch. A fallback that would leave your Mac stops and asks. Leaving the machine is your decision, never Flow's.
Search is a citation, not a guess. 22.3 milliseconds at the 95th percentile on a 10,000-note library, measured 2026-08-04, on Apple silicon. Every result carries an anchor into the exact bytes you wrote. If the passage moved, Flow finds it again. If it is gone, Flow says so.
A fifty-page document is not too big. It is four reviews. Flow plans the file as ordered parts that follow the document's own structure. Each part gets its own review, its own evidence, its own receipt. The finish line is a count a person can check: you approved 3 of 4 parts. Everything else stays exactly as written.
You install Flow. You can still pull a model and run approval-gated work on your own machine. The bundled runtime is about 19 MiB. Two engines. The size of twenty photos. If you already run Ollama or LM Studio, Flow detects them and serves the models you already pulled. Nothing is copied. Your gigabytes are not downloaded twice.
I am not asking you to become a YC partner.
I am asking you to keep the work in the file, and to make the file honest.
A floor of Flow users is a company brain
The talk gets more interesting when you stop picturing one heroic founder.
Picture twelve people who all have their name on documents. Diligence. Comms. Customer work. An operating partner. A research analyst. A founder and the two people who actually write the memos.
In the Roman model, those twelve people become a reporting chain. Someone is paid to collect status. Someone is paid to make the deck that explains the other decks. Intelligence is in the people. The hierarchy routes it. Half of it dies in DMs.
In the Flow model, the folder is the classification. Flow only sees the folders you open. A document cannot reach outside its folder by naming a path. Text changes two ways only: an edit you typed, or a change you approved. Attribution is not a forensics project. It is the product.
Allocation stops being a monthly surprise. Every action carries its own cost record. Local runs record no charge because none was owed. Billed runs record the observed charge. Spend becomes a per-person, per-action fact you can read.
Guardrails live in the document, at creation, as named rules. Not in a policy PDF that nobody opens during the work.
Evidence is scored only where a measurement exists. When the same model writes and judges a change, the receipt says so. You can compare one generation against another without pretending a summary has a quality score.
Routing is a domain decision, not a vendor religion. Long documents are improved part by part, so each part can get a right-sized run.
Curation is a table you can edit in a real grid, while the file underneath stays the Markdown you wrote. Convert Text to Table reports what was actually built, read from the proposed table, never from the model's claim.
The system of record is the files. The search index is disposable. A citation cannot silently drift, because Flow re-checks the passage byte for byte before it highlights anything.
This is the closest I can get, on a Mac, to what the partner meant by making the company legible.
You do not get there by buying a "company brain" t-shirt.
You get there because twelve people did work they already wanted to do, and each approved change left a receipt.
YC's knowledge-mining version of this is thousands of hours of office hours, transcribed, whether the founder liked the camera or not. I will not sell that to a professional services firm. The only capture a professional volunteers for is the one they make themselves, artifact by artifact, because the same record helps them. They consent by name. They get the trust benefit.
The enterprises that win with AI will be the ones that turn real human decisions into validated learning. Flow's receipts are that record. Produced voluntarily. One approved artifact at a time.
When those receipts start to accumulate across a floor of users, the loops can finally talk to each other. Not because a VP routed the information. Because the files are sitting in the same folders, and the next person can see what was approved, what failed, what it cost, and which model did the work.
That is a company brain you can have in Des Moines.
Hill climbing you can actually run
The partner kept returning to a simple test. Did we go up the hill or down the hill. If down, discard. If up, keep.
Flow is opinionated about what "up" means in a document.
Up is a change you can see.
Up is a check that stayed Failed when it failed.
Up is a receipt that names the model, the domain, and the cost, and has nowhere to store your prompt or your API key.
Up is a long document that reports coverage in counts.
Up is a search hit that is still in the file.
Down is a silent rewrite.
Down is a score with no measurement.
Down is a chat session you cannot find on Tuesday.
Down is a fallback that left the Mac without asking.
You do not need a vision document the size of YC's user manual to start hill climbing. You need one outcome you can observe, and a file that will still be there after the model has had its turn.
I have a daily version of this. I call it FOLD, and I wrote it down in the Limitless piece.
Frame the outcome. Orchestrate specialists. Lock the proof. Distill the win.
Flow is that loop as a Mac app.
Frame: the effect sentence, read from the proposed bytes, before you decide.
Orchestrate: five actions, four domains, open models, a frontier checkbox, part-by-part review on a long file.
Lock: the diff, the checks, the approval. Nothing is saved until you say so.
Distill: the receipt stays in the file, so the next run, the next person, the next Tuesday, does not start from zero.
YC's overnight agent files a pull request so yesterday's broken query works today. Flow's receipt is the smaller, more common version of the same idea. The work got better, and the improvement is sitting in the artifact, not in someone's head.
You do not need the badge
I have to say this plainly, because the video will make a lot of people feel late.
You are not late.
You are not behind because you did not get into a batch.
You are not disqualified because your company has Slack DMs, and middle managers, and a policy PDF, and a legal team that will not let you dump every call into a model.
The partner's own caveat is the permission slip. No one knows how to do this. Hundreds of YC companies are trying. The ones who will get there first are the ones who make the work legible and keep a human at the edge.
That is available in a folder of Markdown.
I built Orionfold as one person, one desk, out in the open, after nearly nine years at Amazon. Arena is where I test models. Proof is where a claim has to survive a rerun. Relay is where client work carries a cost receipt. Flow is where that whole family of refusals lands in the document you already write.
I do not write every note anymore. I conduct.
Conduct beautiful documents with AI agency built in. That is the tagline because it is the job. The agents are not miniature employees. They are other paths through the problem. You keep judgment, taste, risk, and the yes.
The Renaissance version of this, from the last essay, still holds. The industrial model specialized the organization until meetings existed so separated functions could explain themselves to one another. The AI-native model lets a person, or a small floor of people, keep the work together.
YC is at the cutting edge of doing this to a famous institution. I am glad they are. Someone has to push the scary version: an AI that reads the applications, picks the interviews, funds the companies, debugs the pitch. He said that end-to-end loop might be possible by the end of 2026, or the first batch of 2027, whether or not they take the PR risk.
You do not need that loop.
You need the one that keeps your name safe on a document.
Flow is here for that.
A week that is not sci-fi
You do not redesign the company on Monday.
Pick one document that already recurs. A weekly brief. A customer update. A diligence memo. A press note. Something that currently burns two to four hours and dies in five tools.
Day one. Do it the old way, but put the working copy in a plain folder of Markdown. No import. The file is the product.
Day two. Run one bounded action. Proofread, or summarize, or turn a section into a table. Read the effect sentence. Look at the diff. Decline it once on purpose, so you can feel that decline changes nothing.
Day three. Turn on the domain you actually mean. Local if the pages should not leave the room. Cloud prepaid if you already pay for a subscription and you are willing to send the text. Watch the receipt. If a fallback would cross a domain, it has to ask.
Day four. If the file is long, let Flow plan it as parts. Approve two. Skip one. Confirm the finish line speaks in counts.
Day five. Hand the same folder to a second person whose name also goes on the work. They should be able to see what you approved, what failed, and what it cost, without a status meeting.
If you are an enterprise AI owner, add one more day. Read five receipts from five people. Ask the questions you already have. Who is spending. Which folders the models can see. Who changed which paragraph. Which check failed and stayed failed. You will know more than last quarter's vendor slide.
Do not begin with the crown jewels.
Do not begin by granting an experimental agent your entire mail history.
Begin with a useful, reversible, slightly boring file.
Boring files are how a company becomes legible. They do not get you a t-shirt. They get you a Tuesday back, and a record you can defend.
What I want you to take from the talk
The YC partner described a future in which intelligence lives in the system and people live at the edge.
I want that future for people who will never be in that room.
The founder who is raising without a famous accelerator.
The knowledge worker who is tired of being the pipe.
The enterprise owner who is being asked to "adopt AI" and cannot answer, yet, which 30 percent is wrong.
You do not need 4,000 hours of office hours. You need the documents you already write, a diff you can see, a gate you still own, and a receipt that stays with the file.
You do not need to burn the org chart this month. You need one DRI per document: the person whose name is on it.
You do not need to ban Slack. You need to stop letting the real work die there.
YC is showing what the cutting edge looks like when a famous institution points its own software team at itself. That is worth watching. It is not a velvet rope.
The techniques are record the work, make it readable, put a policy on what the model may do, put a gate on what may land, keep a human for the calls that matter, and let the system improve from what actually happened.
Flow is those techniques as a Mac app. Patent pending on the revision-scoped, verifiable part. In development. Used daily for real work at my desk. Waitlist open. A freemium subscription is planned. I will not pretend you can download it tonight.
I will pretend, with a straight face, that you do not need a YC badge to become AI-native.
You need a file, a yes, and a record.
Flow is here to help.
The talk is here if you want the source in his voice: Building And Structuring An AI Native Company.
The first essay in this pair is Limitless, Without the Pill.
Start with one useful loop. Keep it in the file. Approve the change. Leave the receipt.
Then see which part of the company wakes up next.
- Orionfold Flow
- AI-native work
- Building in public
Want the playbook?
This is one entry in the build log. The whole AI Native Business book is free, in PDF and EPUB. Just your email.
Conduct your first document with Flow.
Flow is in development. Join for meaningful updates as self-improving intelligence and AI agency move into the document, plus the launch note when Flow is ready.
One email a week until launch, no more.
Top comments (0)