DEV Community: Manav Sehgal

AI-Native, Without the Badge

Manav Sehgal — Tue, 18 Aug 2026 19:15:58 +0000

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.

Get the free book

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.

Join the waitlist →

One email a week until launch, no more.

Limitless, Without the Pill

Manav Sehgal — Mon, 17 Aug 2026 08:35:59 +0000

Going AI-native has felt like the closest real-world version of that fantasy... without the pill, the blackouts, or the Russian loan sharks.

For years, I was fascinated by the idea that the human brain contained vast reserves of unused potential. Books on lateral thinking and Renaissance intelligence made me wonder whether a person could become more creative, analytical, technical, and strategic at the same time. Then Limitless turned that wish into a cinematic fantasy: one pill, and every dormant capability suddenly became available.

Over the past 60 days, while building Orionfold across research, software, products, publishing, design, and operations, I have discovered that AI does not simply help me do more work. It helps me access different modes of thinking, preserve context between them, and combine capabilities that previously lived in separate departments... or separate versions of myself.

This is not a story about replacing humans with agents. It is about what happens when humans learn to conduct them.

It is about becoming more curious, more multidisciplinary, more ambitious, and perhaps a little more Renaissance in how we work.

The future of AI-native humans may not be artificial superintelligence.

It may be ordinary people becoming far more capable versions of themselves.

How going AI-native unlocked different parts of my brain

For years, I wanted a better brain.

Not necessarily a smarter brain. Just one that could remember everything, connect unrelated ideas, switch effortlessly between disciplines, and finish what it started.

A modest request.

At the time, three works shaped how I thought about this.

Edward de Bono's Lateral Thinking taught me that the obvious path is often merely the path our brain has used before.

Michael J. Gelb's How to Think Like Leonardo da Vinci made me wonder whether curiosity, art, science, physical awareness, experimentation, and systems thinking could coexist inside one ordinary human.

Then I watched Limitless.

Bradley Cooper swallowed a transparent pill and went from blocked writer to financial savant, social phenomenon, political force, and owner of several excellent suits.

Naturally, I thought:

Where do I get one?

No productivity app. No morning routine. No second brain composed of 4,700 untagged Notion pages.

Just one pill.

Ideally FDA-approved.

Unfortunately, NZT-48 had several disadvantages: dependency, blackouts, criminals, and a surprisingly high probability of being chased through Manhattan.

AI has so far offered a better risk-reward profile.

Over the past 60 days, I have been building Orionfold as an AI-native business. One person, one desk, multiple products, multiple functions, multiple agents, several local models, and no traditional team.

Something unexpected has happened.

AI has not made me feel as though I am using more of my brain.

It has made me feel as though I can finally use different parts of my brain together.

That distinction matters.

I did not become smarter overnight.

I became more orchestrated.

The 10% dream

In Limitless, Eddie Morra hears a version of the familiar claim that humans access only a fraction of their brains. The film uses "20%," although popular culture usually rounds it down to 10%. (Wikipedia)

It is a wonderful premise.

It is also wrong.

Modern neuroscience does not support the idea that 80% or 90% of the brain sits around waiting for activation. We use the entire brain over time, including while sleeping. The brain represents roughly 2% of body weight but consumes about 20% of the body's energy. Nature is unlikely to maintain that much expensive biological equipment purely for decorative purposes. (MIT McGovern Institute)

But the myth survived because it expressed something emotionally true.

Most of us feel underused.

We know we can be strategic, creative, analytical, empathetic, persuasive, technical, playful, disciplined, and courageous.

Just not on the same Tuesday.

At work, we are usually rewarded for narrowing ourselves.

The engineer becomes more technical. The marketer becomes more marketable. The executive becomes more executive.

The writer learns to write while someone else handles research, design, distribution, analytics, customer feedback, and the small matter of keeping the company solvent.

Specialization created enormous economic value.

It also encouraged us to leave entire modes of thinking at home.

The real limitation was never that 90% of our neurons were asleep.

It was that our attention, time, tools, and organizations made it prohibitively expensive to activate many capabilities together.

We did not have unused brains. We had poorly routed work.

Digging a better hole

Edward de Bono described traditional reasoning as vertical thinking: logical, sequential, and increasingly deep.

It is excellent when you are digging in the correct place.

The problem is that greater effort does not rescue a bad starting point.

De Bono's famous formulation was:

"You cannot dig a hole in a different place by digging the same hole deeper."

Lateral thinking asks us to move sideways.

Introduce a random input. Reverse an assumption. Use provocation. Find a new entry point into the problem.

De Bono distinguished between generating the strange idea and producing movement from it. A provocation was not valuable because it sounded clever. It was valuable if it helped the mind travel somewhere useful.

This is remarkably close to how I now use AI agents.

I do not ask one model to give me "the answer."

I ask different agents to enter the problem from different directions.

One acts as a product strategist. One behaves like a skeptical customer. One researches technical feasibility. One looks for positioning. One attacks the economics. One turns the result into a prototype. Another tries to break it.

The agents are not miniature employees living inside my computer.

They are alternative paths through the problem.

The value is not merely that they work faster.

The value is that they prevent me from digging the same hole with greater enthusiasm.

Leonardo did not stay in his lane

How to Think Like Leonardo da Vinci presents seven principles inspired by Leonardo's life and notebooks: relentless curiosity, learning through experience, refining the senses, embracing uncertainty, balancing art and science, cultivating the body, and recognizing connections between systems.

It is difficult to read this without feeling slightly inadequate.

Leonardo painted the Mona Lisa, studied anatomy, designed machines, investigated water, worked on military engineering, filled notebooks with mirrored writing, and still found time to make the rest of us look undercommitted.

The book's author, Michael Gelb, was himself a professional juggler who performed with the Rolling Stones and Bob Dylan before applying juggling and aikido to learning and leadership.

This is either proof of integrated intelligence or evidence that career advice used to be much more interesting.

What fascinated me about Leonardo was not simply that he knew many things.

It was that he appeared to see through disciplines.

Anatomy informed art. Observation informed engineering. Water became both physical system and visual form.

Questions migrated from one notebook page to another until boundaries between subjects became less important than the underlying patterns connecting them.

Most modern knowledge work operates in reverse.

We separate research from building. Building from marketing. Marketing from sales. Sales from customer success. Customer success from product strategy. Product strategy from financial planning.

We then create meetings so these separated functions can explain themselves to one another.

The Renaissance model integrated the person.

The industrial model specialized the organization.

The AI-native model may allow us to combine the strengths of both.

What happened when I left the organization

At the end of May, I opened Orionfold to the public.

The premise was simple:

One person. One desk. Out in the open.

After nearly nine years at Amazon, I wanted to understand how much useful work one person could produce when the business was designed around AI from the beginning, not retrofitted with a chatbot after the org chart had already hardened. (Meet the builder)

I was not trying to create an automated company in which machines make every decision.

I was trying to create a company in which intelligence can be routed differently.

Orionfold now operates across several connected layers.

Arena is where ideas and models are tested.

Proof turns claims into reproducible evidence.

Relay turns proven capabilities into client workflows with visible cost, margin, and human approval.

Around them sit research, products, developer tools, models, applications, books, websites, and a growing body of field notes. (Explore Orionfold)

In a traditional company, each of these might require a team.

At Orionfold, they began as conversations between me and specialized agents.

One agent worked on strategy. One worked on marketing. One worked on the website. One operated as a research engineer on my local NVIDIA DGX Spark.

I remained responsible for direction, judgment, taste, risk, and deciding what deserved to exist.

The agents expanded the surface area I could explore.

Within one weekend, a paid Arena Field Edition originally planned as a ten-week project was built and shipped. An Advisor product moved from idea to a publicly tested product in roughly 30 hours. One toolbox went through 16 versions in six days. (Read "The Lab That Shipped Itself")

My public build records now include dozens of field notes, more than 176,000 words, over 63,000 lines in the fieldkit codebase, multiple local models, three books, production websites, applications, benchmarks, and reusable tools. (Read "The Glue Tax")

These numbers are not presented as evidence that every line is brilliant.

Some of the lines are almost certainly plotting against me.

The point is that one person can now move among research, engineering, product, publishing, design, operations, and distribution without waiting for every function to become separately staffed.

I have a phrase for this:

I do not write every note anymore. I conduct.

The orchestra inside one person

The conductor metaphor can sound grandiose until you understand what is actually being conducted.

It is not an orchestra of artificial geniuses.

It is an orchestra of partially reliable capabilities.

Research agents are good at breadth but need source discipline.

Coding agents can move quickly but occasionally construct elegant solutions to problems nobody has.

Marketing agents can produce twenty headlines in seconds, including nineteen that sound as though a software company has discovered fire.

Analytical agents are excellent at producing tables.

They are less excellent at knowing whether the table matters.

The human role does not disappear.

It moves upward and inward.

I spend less time generating every intermediate artifact and more time asking:

Is this a real problem?
Is the proposed solution coherent?
What would make it useful?
What feels dishonest?
What is missing?
What should be tested?
Is this good enough to ship?
Does it belong in the world?

These are not leftover tasks after automation.

They are the work.

AI reduces the cost of expressing an idea through multiple forms.

An intuition can become research. Research can become an architecture. The architecture can become code. The code can become a test. The test can become a receipt. The receipt can become a story. The story can become a product. The product can generate new questions.

That loop once required a sequence of departments.

Now it can occur inside one person's working day.

AI does not remove the need for a capable human. It makes human capability more composable.

My brain did not expand. Its switching costs collapsed.

Before AI-native work, changing disciplines was expensive.

To move from strategy to software, I had to load the technical context.

To move from software to storytelling, I had to reconstruct the customer narrative.

To move from storytelling to analysis, I had to find the data, rebuild the assumptions, and remember why any of this mattered.

Every transition incurred cognitive tax.

By the time the brain had loaded the correct application, the day was over.

Agents change this because context can remain active outside my immediate attention.

A research agent can preserve its sources. A coding agent can retain the architecture. A marketing agent can remember positioning. A testing agent can maintain the evaluation harness. A project agent can record decisions, open questions, and next steps.

I can return to a function without rebuilding it from zero.

This is the closest I have come to the Limitless fantasy.

Not instant genius. Not perfect recall. Not the ability to learn Italian during a taxi ride.

Something more useful:

Continuity across modes of thought.

The AI-pilled edge is becoming visible

There is now a small group of people who have moved beyond "using AI."

They are reorganizing their work around it.

They run several agents at once. They package repeated instructions into skills. They maintain persistent context. They connect agents to repositories, browsers, documents, calendars, messaging systems, and local machines.

They do not merely prompt.

They operate systems.

OpenClaw became a visible symbol of this shift: an open-source, locally controlled assistant capable of working through familiar chat interfaces and performing tasks across a user's own devices.What began as a weekend project went viral, accumulated hundreds of thousands of GitHub stars, and ultimately moved into a foundation-backed open-source model. (GitHub)

The interesting part is not the star count.

It is what people wanted.

They did not want another empty chat window.

They wanted an agent that remained available, knew the environment, and could take action.

The same pattern is appearing inside companies.

OpenAI reports that Codex is used across its departments, including legal, finance, recruiting, engineering, and research. In its 2026 analysis, active Codex usage grew more than fivefold during the first half of the year; more than 10% of users were managing at least three concurrent agents weekly, and long-running tasks increased sharply.

Early AI-native startups are also testing flatter operating models. Some are serving large user bases with extremely small teams, while research cited by The Wall Street Journal suggests AI-centric companies may reach similar valuations with fewer employees than traditional peers.

This does not mean every company will consist of one founder and a warm laptop.

It means the minimum efficient size of an organization may be falling.

More importantly, the minimum efficient size of an ambition may be falling.

Projects that once required permission, capital, hiring, and coordination can increasingly begin with one motivated person.

That is the truly disruptive part.

The FOMO is wrong

The current conversation around AI oscillates between two emotional states.

The first is FOMO: everyone else has automated their company. Their agents wake up at 4 a.m., identify a market, build a SaaS product, negotiate cloud credits, publish a launch video, and send the founder a motivational summary before breakfast. You are late.

The second is FUD: AI is useless. It hallucinates. The economics do not work. It will destroy jobs, creativity, education, privacy, democracy, customer support, and possibly the quality of restaurant recommendations.

Both positions contain fragments of truth.

Neither is a useful operating model.

The evidence does not suggest that everyone has figured this out. Macroeconomic productivity gains remain uneven. Daily intensive usage is still concentrated among a minority of workers and businesses. Academic evaluations of long-horizon agents show that even leading systems complete only a minority of complex cross-application tasks reliably.

This is not evidence that the opportunity is over.

It is evidence that the opportunity has barely begun.

You are not late.

Most organizations are still deciding whether employees are allowed to paste a paragraph into a chatbot.

The frontier is not crowded.

It is mostly people wiring things together and discovering which cable starts the fire alarm.

The FUD is also wrong

The strongest argument against AI-native work is not that the models are incapable.

It is that they are capable enough to create damage while still being unreliable.

That is a serious concern.

Agents can expose sensitive information. They can take incorrect actions. They can optimize the stated goal while violating the unstated one.

Persistent agents connected to browsers, messages, files, or payment systems create a much larger security surface than an isolated chatbot. Research into autonomous-agent safety repeatedly emphasizes the need for restricted permissions, isolation, monitoring, and human review.

The response should not be blind optimism.

It should be better engineering.

At Orionfold, I have increasingly organized important work around frozen tests, repeatable inputs, model and configuration identifiers, cost records, explicit approvals, and outputs that can be inspected.

One experiment uses a short configuration hash so a run can be reproduced with the same input and setup. Another local 4B model scored 18 out of 21 on a governed evaluation and refused all nine trick questions. (Read "Same Input, Same Receipt")

The system has also found its own embarrassing mistakes.

In one case, a paid cloud call was presented as though it were free.

In another, a careful refusal was incorrectly classified as a data leak.

Both bugs were corrected and the tests rerun. (Read "The Fix That Changed the Leaderboard")

This is what optimism should look like.

Not "the model is always right."

Not "the model will be right next quarter."

But:

We can build systems in which being wrong becomes visible, correctable, and less likely to recur.

That is how every serious technology matures.

From knowledge worker to Renaissance operator

The industrial knowledge worker is defined by a function.

The AI-native operator is increasingly defined by an outcome.

A founder may spend the morning on architecture, the afternoon on positioning, and the evening reviewing an evaluation. A lawyer may create a research workflow. A designer may build a working prototype. An engineer may test a market before implementing the product. A domain expert may package years of tacit knowledge into a reusable agent.

This does not make every person Leonardo da Vinci.

Leonardo remains inconveniently difficult to benchmark.

But it does revive an older idea: a person can develop across many domains without treating each interest as a distraction from their official identity.

The Renaissance Man was not limitless because he completed infinite tasks.

He was powerful because knowledge moved freely across categories.

AI-native work makes that movement cheaper.

It allows the analytical brain to consult the creative brain. The strategic brain to consult the technical brain. The ambitious brain to consult the skeptical brain.

And occasionally, the enthusiastic founder brain to consult an agent whose sole purpose is to ask: "Are you sure anyone wants this?"

This agent has prevented several masterpieces.

The Seven R's of the AI-Native Renaissance

After 60 days of building this way, I have arrived at seven working laws.

They are not commandments. They are closer to warning labels written after touching the hot surface.

1. Route the work

Do not ask one giant agent to become your entire company.

Separate the modes of thinking. Use one context for research, another for building, another for criticism, another for testing, and another for communication.

Specialization helps agents for the same reason it helps humans: the role clarifies what good work looks like.

The goal is not to simulate an org chart. It is to give the problem multiple intelligent entry points.

2. Retain judgment

Delegate production. Do not delegate responsibility.

The human should retain authority over goals, truth, taste, ethics, risk, and final approval.

Your agent may produce the proposal. You still own what happens when somebody believes it.

3. Require receipts

A polished answer is not proof.

For meaningful claims, retain the source, test, input, output, model, configuration, cost, and decision. This is especially important when the result looks exactly like what you hoped to see.

AI produces confidence cheaply. Receipts make confidence expensive again.

4. Remember what works

Do not restart every conversation from zero.

Capture the decisions. Store successful instructions. Turn recurring context into skills, templates, tests, and reusable components.

The most important improvement is not that the next model becomes smarter. It is that your system becomes less forgetful.

5. Restrict autonomy

Start with the smallest useful loop.

Let the agent research before it publishes. Draft before it sends. Recommend before it purchases. Prepare before it deploys.

Expand autonomy only after the workflow has become observable and reliable.

"Human in the loop" should not mean the human watches helplessly as the loop drives away.

6. Rotate models

Do not build a religion around one provider.

Some tasks need a frontier cloud model. Some need a fast, inexpensive model. Some need local execution because the data should not leave the room. Some need several models so their outputs can be compared.

The best model is not a permanent identity. It is a routing decision.

7. Reuse success

Every successful one-off should leave an asset behind.

A useful prompt becomes a template. A template becomes a skill. A skill becomes a workflow. A workflow becomes a product. A product becomes a platform capability.

This is how AI-native work compounds.

You do not merely finish the task. You improve the machine that finishes future tasks.

Route. Retain. Require. Remember. Restrict. Rotate. Reuse.

The system I follow: FOLD

The Seven R's describe how I think. The system I use each day is simpler.

I call it the FOLD loop.

F, Frame the outcome

Begin with an observable result.

Not "research the market." Instead: "Identify three customer problems, support each with evidence, and recommend one experiment that can be completed this week."

A well-framed outcome gives both the human and the agent something to reject.

Ambiguity feels creative until five agents interpret it differently.

O, Orchestrate specialists

Assign distinct roles. For a new product, I might use:

A researcher to gather evidence
A strategist to identify the wedge
A skeptic to attack the assumptions
A builder to produce the prototype
An evaluator to test it
A storyteller to explain it

The exact roles change. The principle does not.

Different forms of intelligence should collide before the work reaches the customer.

L, Lock the proof

Before expanding the workflow, capture what makes the result trustworthy.

What was the input? Which model ran? What did it cost? Which test passed? Where did it fail? Who approved the output? Can the run be repeated?

This is where a clever demonstration becomes an operating capability.

D, Distill the win

After the work succeeds, package what worked.

Update the instructions. Save the rubric. Create the reusable component. Record the failure mode. Convert the workflow into something the next agent, and the future you, can use without rediscovering it.

Then fold the result into the next loop.

That is the operating idea behind Orionfold. Each successful piece of work should make the next piece easier.

A five-day way to begin

You do not need to redesign your company tomorrow.

Choose one recurring piece of work that currently consumes two to four hours.

On day one, write down how you perform it today. On day two, ask an agent to complete one bounded part of it. On day three, define a simple rubric for evaluating the result. On day four, separate production from review: one agent produces, another critiques, and you approve. On day five, save the successful instructions, examples, and rubric as a reusable workflow.

Do not begin with your most sensitive data. Do not begin with financial transfers. Do not begin by granting an experimental agent access to every message you have written since college.

Begin with something useful, reversible, and slightly boring.

Boring workflows are excellent teachers. They do not generate viral demos. They generate saved Tuesdays.

What "Limitless" means now

I no longer want to use 100% of my brain. That sounds exhausting.

I want to use the correct mode of thinking at the correct moment, supported by systems that preserve context, generate alternatives, test claims, and turn successful work into reusable capability.

That is what going AI-native has begun to unlock for me.

Not intelligence from nowhere. Not instant expertise. Not a pharmaceutical montage.

A more connected version of the person already there.

The strategist can work with the engineer. The engineer can work with the writer. The writer can work with the researcher. The researcher can work with the entrepreneur. The entrepreneur can work with the critic.

And all of them can finally leave notes for one another.

The future of AI-native humans will not be humans doing nothing while machines run the world.

It will be humans with greater agency.

A teacher who can create a personalized learning system. A consultant who can turn expertise into software. A scientist who can explore more hypotheses. A small business owner who can finally afford capabilities previously reserved for large companies. A founder who can test ten ideas before hiring for one. A curious person who no longer has to choose a single lane before being allowed to begin.

Some people will use AI to produce more noise.

Others will use it to create more leverage, more independence, more experimentation, and more ambitious forms of work.

The distinction will not come from access to the model. We will all have models.

It will come from the systems we build around them, and the judgment we retain within ourselves.

I once wished for a pill that would unlock the rest of my brain.

It turns out I did not need another 90%.

I needed a better way to coordinate the 100% I already had.

No NZT. No Russian loan shark. No side effects observed so far, other than an unreasonable number of repositories.

I am publishing the experiments, products, receipts, failures, and next folds as I build them in Orionfold Stories.

You do not need to follow my exact path.

Start with one useful loop. Route the work. Retain the judgment. Require the receipt.

Then see which part of you wakes up next.

This essay first appeared on Orionfold Stories. Orionfold Flow, the AI agency for documents on Mac, is in private waitlist: join it here.

Same input, same receipt

Manav Sehgal — Sun, 16 Aug 2026 19:28:05 +0000

Most AI benchmarks you read are screenshots. A number, a chart, a confident sentence about which model won. You cannot rerun them. You cannot change one input and watch the result move. You have to take the author's word for it, and the author had every reason to pick the run that looked best.

I did not want to ship another tool that produces screenshots. I wanted the opposite: a result you can pick up, rerun, and check. So every receipt Orionfold Proof writes carries a small string near the bottom, and that string is the whole idea.

The twelve characters

Here is the line, from a real receipt, the four-billion-parameter model that scored 18 out of 21 on my laptop:

Run id:      run_a9014d0871ab
Config hash: 50c38b0b7439   (identical inputs reproduce this hash)
Rerun:       POST /api/runs { "dataset_id": "advisor-curveball-v0.2",
             "candidate_ids": ["ollama:hf.co/Orionfold/Advisor-GGUF"], ... }

That 50c38b0b7439 is the config hash. It is computed from everything that determines the result: the dataset, the exact examples, the candidates, the prompts, the scoring rubric. Feed the same inputs in, and you get the same twelve characters back, every time. The receipt even tells you how to reproduce it. The Rerun line is the literal request that made it.

This is not a checksum of the output. It is a fingerprint of the setup. Which means two people on two different machines, who have never spoken, can run the same proof and confirm they ran the same thing, because the hash matches. The result becomes a fact you can check, not a claim you have to trust.

Change one input, change the hash

The hash is only useful because it is honest about change. Touch anything that matters, and it moves.

Swap the model, and the hash changes: you are proving a different thing. Edit one example in the dataset, and the hash changes: your test is no longer the test you cited. Loosen the rubric threshold to make a model look better, and the hash changes, and anyone holding the old receipt can see you changed the question.

That last one is the quiet protection. The most common way benchmarks lie is not a fabricated number; it is a moved goalpost. Run it, do not like the result, nudge the threshold, run again, publish the good one. With a config hash, the nudge is visible. The receipt that says 86% at one hash and the receipt that says 86% at a different hash are not the same claim, and the difference is right there in twelve characters.

I caught this on myself last week. I ran the same governance bench on the same three models twice. The first time I forgot to give the models their instructions, and all three scored a dismal 1 out of 21, at hash b6fa4ce299ef. The second time I supplied the contract, and they scored 18, 17, and 16, at hash 88403c2fc4e7. Two different hashes, two different results, no ambiguity about which run was which. The hash did not let me quietly pretend the first run never happened, or that the second one was the only one. If I had only shown you the good numbers, the hash on the receipt would not have matched the bench I claimed to run. The fingerprint keeps me honest even when I would rather not be.

Why a receipt, and not a dashboard

There is a deeper reason this matters, and it goes back to who the tool is for.

The person using Proof is usually deciding what to trust for someone else: a client, a team, a future version of themselves who will have forgotten the details. A dashboard is for watching. A receipt is for handing over. It is the artifact you attach to a proposal, the thing you can show a client under NDA without leaking a key, the record you can pull up in six months and rerun to check it still holds.

So the receipt is the product, and the hash is what makes it a receipt instead of a screenshot. It is signed. It is rerunnable. It records the failures alongside the wins, on purpose, because a record that only flatters is the one a careful buyer throws away.

The honest part

A config hash does not make a result correct. It makes it reproducible, which is a different and smaller promise. A reproducible benchmark can still ask a bad question, score it with a bad rubric, or measure the wrong thing entirely. The hash only guarantees that if you run what I ran, you will see what I saw.

But that smaller promise is the one almost nobody else makes, and it is the floor everything else stands on. You cannot argue about whether a benchmark measures the right thing until you can agree on what was run. The hash gets you to that line. After that, the argument is about the question, which is the argument worth having.

That is the whole philosophy of the thing: prove it yourself, on your own machine, and rerun the proof anytime. You can see what it produces, including the receipts behind the other stories, at orionfold.com/proof.

Originally published at orionfold.com.