Frame width
The large number is the true width of what fills your screen, in metres, given in scientific notation and again in the largest sensible unit. It is interpolated continuously as you scroll, so it is never rounded into a convenient lie.
David HarderFractional AI Director
Orders
of Magnitude
One continuous fall from the observable universe down to a single proton. Forty-one orders of magnitude across thirty-two stations, every one of them drawn the moment you ask for it. At the exact middle of the descent there is a man on a blanket in the late afternoon, and the man is me.
An exhibit, and also an argument. You are not too small to matter, and not too busy to change. You are standing at the one point of leverage in your whole operation, and my work is finding it.
Begin the descent What I actually do
Scroll slowly, the instrument keeps the count beside you. Or skip the poetry and go straight to the work.
00
How to read the instrument
The rail
Forty-three ticks, one for each power of ten from 1027 down to 10-15. The playhead marks where you are. The coloured thread beside it is the palette of the whole journey: cold at the top, warm through the middle, electric at the floor.
True scale
Wherever a real measurement exists, it is used. The Moon sits 384,400 kilometres out. The helix is two nanometres wide with one full turn every 3.4. A proton's charge radius is 0.84 femtometres. Sizes are computed from the frame width, never eyeballed.
One zoom, no cuts
Each station is drawn at ten to the power of its own width minus the current width. Worlds therefore grow through one another rather than replacing one another. If you cannot name the instant the cosmos became a galaxy, that is the whole exhibit.
The descent, thirty-two stations from the observable universe to a single proton
This section is a scroll-driven animation. Every station also appears, as a still plate with the same caption, in the atlas below.
The scroll-driven descent is paused. Reduced motion is switched on, or you have chosen atlas mode. The full journey is below as thirty-two still plates, each drawn by the same code. Go to the atlas
33
What just happened
You fell forty-one orders of magnitude, and on that ruler you are almost exactly in the middle. A human being is as far beneath the observable universe as a proton is beneath a human being. Nobody arranged that. It is a fact about where creatures like us can happen: large enough for chemistry to get complicated, small enough for a planet to carry us.
Twice on the way down the frame went nearly empty, and both times it was telling the truth. Between the Earth and the Moon there is almost nothing. Inside a single atom there is even less. Matter is not dense. It is a rumour held in place by fields, and the tiny fraction of it that is anything at all is mostly the energy of its own arguing.
The middle of the journey is the strangest part. Somewhere around one metre the numbers stop being astronomical and start being personal, and for about four orders of magnitude the exhibit is simply a person on a blanket in a park in the late afternoon. Everything above that frame is too large to touch you. Everything below it is too small to notice you. You live in the only band where both directions are visible at once.
Charles and Ray Eames made Powers of Ten in 1977: nine minutes of film that began on a picnic blanket in Chicago and travelled this road in both directions, building on Kees Boeke's 1957 book Cosmic View. This is a browser's homage. Same argument, different instrument, no camera anywhere in the building.
So the middle of the ruler is the interesting place to stand. Everything below you is mechanism. Everything above you is consequence. That is also true of a business, which is the part of this I get paid for.
34
The man on the blanket
I'm David Harder. I build agentic systems for people who run real businesses, and I stay accountable for them after they are built.
Here is the gap I keep finding. Most executives have met AI exactly once, as a better search box. You typed a question. You got a decent paragraph. You were mildly impressed. Then Monday happened anyway, and nothing about your week had actually changed.
That is not the work.
The work is systems that do the job while you sleep, with a human gate exactly where it matters. Not ask the AI about your invoices. The invoices get chased, you approve the exceptions from your phone, and the four hours come back. You stop being the bottleneck in your own company.
I am not going to teach you to prompt. I will take three things off your desk and stay accountable for whether they stay off.
The role has a name and I would rather say it plainly. Fractional AI Director. Director, because somebody has to be accountable and hold the seat. Fractional, because you do not need another full-time hire, you need the thing done.
As for why you have just fallen through the universe to get here. Anyone can buy a consultant's template. This page is drawn frame by frame in code, thirty-one of its thirty-two plates contain no photograph at all, and it was built by a fleet of AI agents I directed rather than a studio I hired. It is the only credential I could think of that you cannot buy and cannot fake. You are standing inside the demonstration.
35
The work
Three systems that are running right now. Each one is written the same way: what the situation was, what an agent took over, and what it gave back. The failures are in here too, on purpose.
Marketplace & payments
The marketplace that pays thirty makers correctly
A Canadian artisan marketplace needed a commerce spine that could actually take money and split it right. The hard part was never the shopping cart. Thirty small businesses each have a different tax position, and getting that wrong is the founder's legal problem, not the developer's.
- What an agent owns
- Stripe Connect payouts per maker. GST charged only when a maker's own registration says to charge it. Per-maker shipping. Referral credit resolved across three seller tiers. And the founder makes changes by texting one word from her phone: an agent reads the request against a written playbook, makes in-scope edits, ships them to production, and verifies them on the live page before she gets a reply.
- What stays human
- Anything touching money, security or contract scope stops. It becomes an approval job and waits.
- What it gave back
- Contracted at one tier, delivered nearer the tier above it. Now on a monthly hosting agreement. A payout failed once because a completion flag in our own database said a maker's account was ready and the payment provider had quietly disabled it. That lesson is written into the operating notes, and the next marketplace will not have that bug.
Regulated content
The practice that stopped paying an agency
A counselling practice was paying an agency to run its social presence. Content in a regulated field cannot claim outcomes, cannot promise timelines, and cannot put an honorific in front of a clinician whose registration does not carry it. An agency gets that wrong occasionally. A regulator does not grade on effort.
- What an agent owns
- Twice a week it drafts the next scheduled post, with the compliance rules written into the generation itself rather than checked afterwards. It emails the draft for approval. A clear yes publishes to both channels and writes an audit row. On Mondays it also writes, commits and deploys a blog post.
- What stays human
- The approval. Every time. Changes requested means revise and resend in the same thread. Anything ambiguous holds and pings me.
- What it gave back
- The agency contract ended and posting moved in house without anyone in house doing it.
Research & analysis
The panel that killed a $282 million number
A client asked for business plans built by agents argued against each other rather than agreeing with each other. The failure mode in AI-written plans is never the prose. It is a confident number with nothing underneath it, arriving in front of a lender.
- What an agent owns
- Research agents work the web and the client's own source documents. Then a five-lens adversarial panel: a bull, a bear, a lender, an operator and a protocol reviewer. Claims the panel kills never reach the writers at all. Section writers then produce the document, and an editor makes a consistency pass.
- What stays human
- Every claim that could not be substantiated carries a visible requires-validation marker instead of being quietly smoothed over. The finished document ends with a report on exactly how it was made, so the reader can audit the machinery.
- What it gave back
- One run: nineteen agents, twenty-three minutes, eighty-four findings, one hundred and eighty-seven citations, twenty sections. Twelve claims were killed before a writer ever saw them. One of them was a $282.2 million headline figure that quietly included a scope the project had already cut. The panel also surfaced a 2025 court ruling that made an Aboriginal title declaration a precondition on the entire land strategy, which no research agent working alone had raised.
Client confidential. Numbers above are the process, not the plan.
Or just click something and watch it run
The rarest thing I can offer is not a deck. It is a working system with a URL.
- The FABLE collectionThirty-one production websites built, critiqued and deployed by AI agents over one weekend. Every one carries its own build guide and an honest score from a separate critic agent, including the low ones.
- NESHE NetworkA live marketplace moving real money to independent makers.
- Clearview CounsellingA regulated practice whose content pipeline runs itself, under approval.
- Perrin Mental Health ServicesA clinical practice site built and maintained end to end.
- Syntac clinical intakesA conversational intake engine for therapists, on a database hardened for health information: row-level scoping, an audit trail, and a de-identification policy.
- SynapseMy own operator cockpit. Email, WhatsApp and iMessage unified into one person-shaped view, with drafts staged and nothing sent without a tap. Gated, but it is the demo I give on a call.
36
What I will and will not automate
Most of this industry sells you the first column and stays quiet about the other two. The other two are the reason it works.
An agent can own this
- Drafting, in your voice, learned from what you have already sent
- Triage: what needs you today, what is waiting on somebody else, what is just noise
- Scheduled publishing, after a human yes
- Filing, sorting and organising anything at volume
- First-pass research and analysis, with its sources shown
- Shipping in-scope changes and verifying them on the live page
- Watching for things going wrong and escalating them to a person
This stays human
- The send. Anything going to your client, in your name
- Money moving, in any direction
- Changes to scope, price or contract
- Any clinical, legal or financial claim
- Hiring, firing, and anything about a person's job
- Anything a regulator would grade
- The apology, when something has gone wrong
I will not build this
- An AI that presents itself as a person to someone who has not been told
- Reviews, testimonials or case studies that did not happen
- Your customers' health or financial records into a system that trains on them
- Outreach at volume to people who did not ask for it
- Anything designed to replace a person before that person has been told
- Claims about results I have not measured
On the record
The time it went wrong
In July one of these loops sent the same approver four near identical emails in a day. Two bugs compounding: a handler retried its own command three times, and the reply checker only ever read the newest message, so an approval was thrown away when a question arrived three minutes behind it.
She is a working clinician with a caseload. One retry storm spends trust that took months to earn.
So the fix was not a patch. Every outbound message on that lane now passes through a guard that no code path can bypass: a master hold switch, a forty-five minute cooldown between messages to any approver, and a ledger keyed to each individual reply so nothing is ever acted on twice. The guard keeps its state outside the database on purpose, so it still works when the database does not. Both scheduled jobs were switched off and left off until a human turned them back on by hand.
I put this on my own website because a practice that sells trust and only publishes its wins is selling the opposite of trust. Your systems will misbehave eventually. What matters is what was already built for the day they do.
37
Three ways in
The first one is meant to be an easy yes, and you keep it either way.
01
The Descent
Two weeks, fixed scope. I map where your time actually goes, then hand back a ranked list: what an agent can own, what has to stay human, and what nobody should automate at all. Written plainly, with the reasoning shown.
You keep the map whether or not you ever hire me again.
Fixed fee, low four figures
02
The Build
I build and ship the top two things on your map. Real systems, in your stack, with your approval gates, deployed and running. You watch it happen rather than receiving a report about it.
Scoped from the map, so you know the shape before you commit.
Per engagement, mid five figures
03
The Floor
Ongoing fractional direction. I run an agent floor for your business month over month, keep it healthy, extend it as you grow, and I am the one accountable when it misbehaves.
A seat, not a subscription. Cancel with a month's notice.
Monthly retainer
Who this is for
Not startups. Startups will do it themselves, badly, and enjoy it. This is for the owner or the executive of an established business doing real revenue, who knows this matters, is quietly embarrassed at how little has been done about it, has no time to learn, and would rather not be sold to by someone half their age with a growth deck.
If that is roughly the week you have been having, the first step is small and you keep the map.
Start with the map Straight to my inbox. A real reply from a real person, usually inside a day.
38
The atlas
Thirty-two plates, each one rendered from the same procedural code that draws the descent, held still so it can be read. Frame widths are true. Choose any plate to fall from there.
39
The ledger
Every station, its zone, and the real width of its frame. This table is the engine's own data: the drawing code reads these numbers, so what you see and what you read cannot disagree.
| No. | Station | Zone | Frame width | In words |
|---|
40
Colophon
- Drawing
- Canvas 2D, entirely procedural. Ridged multifractal noise with domain warping for the cosmic web, particle spirals for the galaxies, real orbital radii for the solar system, midpoint relief for the coastline, recursive subdivision for the city grid, bezier anatomy for the figure and the hand, parametric geometry for the double helix, rejection-sampled orbital probability for the electron cloud.
- Motion
- GSAP with ScrollTrigger, scrubbing one normalised journey timeline. The zoom itself is arithmetic: ten to the power of a difference of logarithms.
- Type
- Newsreader for the wonder, IBM Plex Mono for the data. Nothing else.
- Photographs
- One image, appearing at stations twenty through twenty-two. It is a generated image, built from my own headshot with my consent, and it is honest about being one: same man, same build, a picture that was made rather than taken. It will be reshot for real, and when it is, this line will say so. Everything else on the page, all thirty-one other plates, is drawn by code at the moment you ask for it.
- Authorship
- The exhibit engine was designed and written autonomously by Claude Opus 5, working from a brief and an iteration loop I directed. The words, the argument and the positions in the trust ledger are mine. I would rather say which is which than let you assume.