A proposal · the age of agents
Here is a thought — not written in English, not in any tongue anyone speaks, but in the writing system this page is about:
Do you know what it means?
It says: "Let us calculate, to see who is right." — the words of G. W. Leibniz in 1685. And it is the only way to write that thought — two people would type it identically, and so would two machines.
Now, here is a more complex thought — the opening line of Alan Turing's 1950 paper on AI:
It says: "I propose to consider the question, 'Can machines think?'" No dialect, no drift, nothing lost between them. One meaning, one form.
read on
And it is not only for new sentences. Anything ever written has one exact form. Here are two famous lines — first in the new language, then in the words you know:
Man can do what he wills but he cannot will what he wills.
Three minds, centuries apart, reaching for the same thing — a way to write meaning so exactly that nothing is lost passing it on. This is not a historical recreation, but the realization of the concept Leibniz envisioned 350 years ago. He was an amazing visionary, and now that we finally have the engine (LLMs) and the urgent need (AI coordination), it is time for his vision to come alive.
The stakes
We are entering a workday where each person oversees not one model but millions of agents delegating, summarizing, and acting on one another’s words. Picture the traffic: millions of hand-offs per person, every one a chance to lose what you meant. We are about to scale our communication with machines a millionfold — and the channel we are scaling is the leakiest one we have: natural language, built for humans who share a world the machine has never seen.
Every age that mattered first learned to write down the one thing it could not afford to lose. The alphabet captured speech; numerals captured quantity. Before the 11th century, if you wrote a song in Italy, a musician in China could never play it. But then humanity invented staff notation — a universal visual mathematics for sound. Suddenly, a composer in Germany could write a symphony, hand it to a musician in Japan 300 years later, and it would be played flawlessly.
The age of agents runs on something we still cannot write down exactly: meaning. What we want. What we decided. What was meant. Right now, if an English doctor writes a medical prescription, a Chinese pharmacist needs a translation app, an insurance database needs an API, and an AI agent needs to parse the fuzzy text — and at every hop, a mistranslation could be fatal. But if that prescription is written in Alan, the doctor, the pharmacist, and the machine are all looking at the exact same immutable logical graph. No translation is needed. Alan is the staff notation for human thought.
Why now
To stay genuinely in the loop — to steer and catch the errors instead of rubber-stamping — a human has to grasp exactly what the machine means. But words are too loose; they were made for people who already share a world. Code is too low; it says what to do, never what was meant. Nothing sits in between. The collaboration of the age has no notation.
The damage, measured
This is not speculative. Every seam where meaning crosses into or between machines is already measurably failing:
Every one of these is the same wound: meaning poured into a medium that cannot hold it, then carried between systems that share no ground. Multiply it by a million agents and it stops being a bug list.
It is a wall.
The workaround
The evidence is in your repos. Developers now commit prompts to git — paragraphs of English, versioned and diffed and traded like source code — because nothing better belongs in between. And every attempt to firm it up repeats the same mistake: wrap the prompt in XML, in JSON, in a “structured” template. That is only structure — boxes around the same ambiguous words. A schema is not a writing system. What’s missing is a notation whose structure is logic itself, where the shape of what you write is the meaning.
The fix
So stop patching words and write the meaning down directly — in one universal notation shared by everyone at the table, human and machine alike. One rule carries the whole thing: one meaning, one form. Each symbol means exactly one thing; each thought has exactly one spelling. Re-write it and you have written the same thing — nothing to paraphrase, nothing to blur. Mathematicians call an operation that changes nothing a fixed point: T(x) = x. Pass meaning through one mind or a million, human or machine, and it arrives identical.
A writing system of pure meaning, for the age we and the machines will write together.
You would never write it by hand. You keep your English; the model compiles it down into the notation and renders it back — the one record in the middle that both sides read, verify, and trust. There is no magic in the economics, and that is the point: pinning a meaning costs a fixed number of bits. Natural language hides them in shared context the machine does not have, so the machine guesses. This is the only design that makes the writer pay those bits once, in the open — instead of making a million readers guess.
The benefit to AI
Give every party one form, and each crack above loses its cause:
T(x) = x, so nothing mutates between hops, and invariance is checkable at each seam — inter-agent misalignment has nowhere to enter.Notice what happened to the translation layer everyone was paying for. It did not get more accurate.
It disappeared.
See it in motion
An instruction re-summarized at every hop frays; the same instruction re-emitted in canonical form is the identity — T(x)=x — so it arrives untouched. Watch:
English — re-summarized
A: “Deactivate accounts inactive since last spring, except admins.”
B: “Disable old unused accounts; skip admins.”
C: “Clean up stale accounts.” ✗ admins lost ✗ date lost
meaning frays at every hop
The notation — re-emitted
B: ... identical logical graph ...
C: ... identical logical graph ...
✓ T(x) = x — nothing lost
It was found, not invented
None of this is invented; ideas like it are found. In the 1670s Gottfried Leibniz — who co-discovered calculus — drew exactly this, in two parts. A characteristica universalis: symbols that stand for concepts, not sounds. And a calculus ratiocinator: an engine to compute and check over those symbols. Get the pairing right, because it is the entire point — the symbols hold the meaning; the engine reasons over it. Settle a dispute, he said, not by arguing but by calculating — calculemus.1
The dream was never the bottleneck — the engine was.
For three hundred and fifty years it sat in a drawer. The engine now exists, and it is the thing you already use every day: a large language model is the calculus ratiocinator — the bidirectional translator that compiles your fuzzy English into the form and renders it back to be checked. The characteristica is the canonical form; the model is the calculus. Together they make meaning computable and checkable — the only way anything, human or machine, has ever earned trust.
And the pieces are real, not wished-for. Linguists spent fifty years finding the ~65 semantic primes every language already shares2; Abstract Meaning Representation already proves one meaning can map to one graph3; UNL and the controlled languages built the language-neutral scaffolding.4,5 The need is screaming, too: the frontier labs are already letting models reason and talk in private codes of their own — continuous latent thought14, byte-patch internals15, agents trading raw hidden vectors for a 24× speed-up.16 The instinct is right; the form is backwards. A writing system you can read beats a cipher you can’t. Revived ideas come back wearing the substrate’s clothes — Leibniz’s characters return not as ink, but as the notation the machines pass between themselves.
The Complete Dictionary
About sixty-five symbols for the handful of meanings every language already shares — the atoms — plus a grammar of spatial connections and logic signs from math. Sixty-five primes, the way a hundred elements give you all of chemistry: coverage comes from a small generative basis, not an infinite dictionary. You could learn them in an afternoon.
We have fully realized these 65 primes into a comprehensive visual grammar, categorized by their geometric families (Entities, Actions, Time, Space, Logic).
Why it happens
Nothing about it asks the world to wait. No one changes how they talk — the work lives in the model, the one part of the stack that gets smarter on its own every month. Two agents that share the notation stop leaking today, and every new one that learns it makes the whole network worth more — the flywheel that carried the alphabet, the numeral, and Unicode from a handful of users to everyone. With millions of agents arriving17, that flywheel is the steepest in the history of computing. Every age agrees, eventually, on how to write down its core act. The core act of this age is passing meaning between minds — and it has none.
This is that notation.
The first clean page
Every age wrote down what it refused to lose: speech, number, music. This age — the first we share with minds of our own making, soon a million of them to a person — will write down meaning. The sketch is three hundred and fifty years old; the engine is two. For the first time, both exist at once, and the machines are already at the table, already building, already waiting for a page they can read.
This is the writing system for the age of AI.
Sources