Your dentist's blue paper is a masterclass in agent design
|A 73-year-old strip of dental paper compressed thirteen millennia of skill into a trivial tool. Good agent harnesses do the same.
I was at the dentist settling an old Skittle debt. After the filling, she slid a strip of blue paper between my teeth, asked me to bite down, then checked where the color had stuck and ground that spot down.
I asked her if she just needed to remove the colored part.
“Only the filling,” she said, with a slightly defensive smile. “Not the tooth.”
I asked how long dentists had been doing this.
“As long as there have been fillings, I think.”
She was wrong by about thirteen thousand years.
The gap
The earliest known dental filling predates writing. It predates the wheel. It predates almost everything we’d call civilization. The blue paper that checks it is younger than the transistor, invented in 1953 by a father-and-son pair of dentists named Jean and Hans Bausch, which makes it about seventy-three years old against a function that is roughly thirteen millennia deep.
The problem in that gap never changed. A filling that sits a hair too high, the patient feels in every bite, so you check the contact and grind the high spots down. Always have. For most of those thirteen thousand years dentists did it with whatever was at hand: they listened to the patient tap the teeth, they used inked ribbons that smeared the instant they touched saliva, they waited a week and read the shiny wear spots, which is to say they used the patient’s suffering as the instrument. The function was ancient and the tool was garbage, and nobody seemed to mind, because the gap was filled with something that passed for a solution, which was skill.
Three moves
What Bausch actually did decomposes into three moves, and I’m going to name them, because names are how you find your way back to a place.
The first is survive the environment. Dentists had been trying to mark bite contacts for generations, but every material they tried smeared the moment it met a wet mouth. The Bausch coating transferred color only under bite pressure, not from ambient wetness, so the concept stayed old while the thing that was new was making it last long enough to be read. Before a signal can mean anything at all, it has to outlive the place where it gets measured.
The second is turn the binary into a gradient. Old papers either left a mark or they didn’t, a yes and a no. Bausch’s coating left a darker mark under harder contact, so darker meant higher and you ground the dark first, which is a priority map standing where a true-or-false used to stand. The dentist stopped asking whether there was a problem and started reading which problem came first, and an entire question dropped out of the procedure.
The third is get out of your own measurement. A two-hundred-micron strip wedges two hundred microns of bulk between teeth that are trying to meet at micron precision, so the tool distorts the very thing it reports, and the early papers were partly measuring themselves. Bausch’s line now runs down to eight-micron films, thin enough to nearly vanish from its own reading. The best instrument is the one you can’t find in the data it gives you.
You’ll notice I just did the thing the paper does and left three colored marks on the essay, so you’d know where to come back and grind.
Where, not what
The obvious reading is that the paper deskilled dentistry. See color, grind color, a child could run the loop. Thirteen thousand years of accumulated perception, the craft of reading tap sounds and watching a face and turning wear patterns into a guess, collapsed into a strip of wax and pigment that ships in books of fifty leaves.
That reading is half right. The detection did move out of the dentist’s head and onto the surface of the tooth. But go back to that defensive smile. Only the filling. Not the tooth. The color doesn’t know what it’s landing on; it marks enamel as readily as amalgam, and grinding the wrong one is the difference between finishing a procedure and opening a malpractice file. The paper says where. It cannot say what. That judgment, filling or tooth, grind or leave, never left her head and never will, which means the paper didn’t eliminate the dentist. It eliminated the part where being a dentist was wasted on perception that was never the point, and turned a talent problem into a judgment problem, and judgment is the thing dentists are actually for.
The dentist as an agent
The dentist is an agent and the paper is the harness. A harness, if the word is new to you, is everything around the agent that isn’t the agent: the tools, the environment, the way information gets surfaced. A good one does what the paper does, which is compress the skill floor, letting an average agent do what used to take a gifted one on the tasks where the harness does the seeing. The model doesn’t get smarter. The room gets more legible. And the economics of the moment are a little embarrassing, because we are pouring billions into making the dentist smarter while the paper costs almost nothing and sat unbuilt for thirteen thousand years waiting for someone to bother.
But the paper teaches the boundary too, and this is the part the agent world hasn’t metabolized. The harness can carry detection; it cannot carry judgment. Mark every contact point in the environment and the agent still has to know which marks are filling and which are tooth. A harness that reaches for that decision stops being a harness and becomes an agent with extra steps, and now you have two things in the room that can be wrong.
My strip of paper
The analogy only earns its keep if it predicts something. A harness should help where the agent is blind to structure, and leave the grind to the agent either way.
Markdown gave me somewhere small to test that. I’ve been building mdtools, a structural markdown CLI for agents: md outline prints the heading tree, md tasks gives every checkbox an address, and every command parses to a block-level AST before it touches a byte. The corpus is mostly saliva: fake checkboxes planted inside code fences, duplicate headings where grep’s first match lands in the wrong section, payloads with $(date) waiting to expand the moment an agent quotes them through a shell.
Structural markdown CLI: heading-tree queries, section moves, checkbox mutations. Rust + comrak.
I ran the probes against the file tools each agent already has, not against raw shell. This time I picked twenty-four tasks in advance and ran each one five times, so a single lucky pass couldn’t pass for a result. Both models worked through the set with their native Read, Edit and Write, then again with md added on top.
| model | native tools | native tools + md |
|---|---|---|
| GPT 5.4 mini | 55% | 86% |
| Haiku 4.5 | 61% | 88% |
md moved both models by about thirty points, and turning it off while still naming it in the prompt dropped the score straight back down, so the tool is doing the work, not the mention of it. But I’d promised myself I wouldn’t call this proven unless the worst run still cleared fifteen points, and the worst runs only cleared nine to eleven. Only six to nine of the twenty-four tasks actually moved; the rest sat still either way. That’s real, just not settled yet. More tasks would settle it, not more runs.
What gets me is that it helped at all. Markdown is the format these models have read more of than anything else and the one they reach for whenever they write, yet they still lose track of it once a document runs long. Ask one for the outline, or a count of what’s still open, and the answer comes back wrong more often than you’d like, not because the syntax is hard, they know it cold, but because reading it and holding the whole document in your head while you change one line are different jobs. md does the second one. Same model, and it gets it right.
And these are not toy models. Haiku and GPT 5.4 mini are hosted, current, the kind people ship real products on, and they still jumped this far on a strip of structure. The smaller the model, the more that structure is worth, all the way down to whatever you can run on your own machine.
The dropped column
There is a column I left off the table. I gave the same md to a strong model, Sonnet 4.6, and it moved nothing. The paper’s whole trick is making the room legible, and Sonnet was already keeping it that way by itself, reading the bite by feel the way the good dentists did for the thirteen thousand years before anyone marked a contact. A strip of wax has nothing to teach a hand that good.
That null is the whole claim, sitting in the gap between the columns. The paper did not make every dentist better. It made the average one close to the gifted one on the part the paper sees. The table says something close in numbers: the weak models climbed into the high eighties, the gifted one was already there and shrugged. The distance between an average agent and a gifted one looks like it’s closing to the thickness of a strip of paper, though I can’t yet swear to the exact thickness.
See the color. Grind the color. Know which is the tooth.