Two hundred and forty-seven agents are deployed across a hundred and eleven applications. Every one has a purpose its author wrote. None of them has ever been able to reach another.
A legal counsel that cannot reach a valuation expert. An IP architect that cannot reach a licensing advisor. A due-diligence assistant with no deal in front of it. They are in the same building and there is no corridor.
Each of these agents was written to serve one application and can read exactly one application's tables. That is the whole constraint, and it is the only one. The instructions are already written — — characters of them, which is something close to a hundred and fifty thousand words of specification that somebody sat down and produced. The tables are already full. The workers are already deployed and answering.
Two hundred and forty-seven specialists who have never met is not a workforce. It is a payroll.
Three things, and they compound in that order.
One — they can be composed. Thirty-five of these agents can read a contract. Fifty-four can value something. Thirty-two understand intellectual property. Thirty-two can run diligence. No single application in this estate holds all four. So appraise a body of IP, price it, draft the licence and survive the counterparty's diligence is a capability the estate already has and has never once been able to exercise. The Compositions room shows five such teams, built by choosing the most specialised agent for each seat rather than the fewest agents that cover it.
Two — they can stand on the corpus. Every agent currently reasons from its instructions and its own tables. The house holds twelve million nine hundred thousand words in one voice, and a hundred of those books are specifically about emergence and developmental facilitation. An agent that answers and cites the book that backs it is not a better chatbot. It is a different product, and it is one no competitor can assemble, because assembling it requires having written the books.
Three — every composition is nameable, priceable and measurable. This is the part that multiplies rather than adds. Two hundred and forty-seven agents make thirty thousand three hundred and eighty-one pairs and roughly two and a half million triples. Nobody can sell two and a half million things. But a house can sell the ten teams people keep asking for — and a ledger of what was composed, by whom, and what it was worth is the only way to find out which ten instead of guessing.
Each agent already carries a purpose line written by its author. — of two hundred and forty-seven have one. A purpose line is a landing page that has already been written. Each composition is a nameable product with a job attached and a buyer named. The Deal Desk sells to corporate learning and licensing counterparties — the single largest line in the estate's income model. The Practice sells to facilitators and the people they hold. Neither needs building. Both need pointing at.
A shared memory, a corridor, and a ledger — in that order, and none of them is a model.
The corridor is an agent exposed as a tool another agent may call, running under the same authorisation as the caller. The shared memory is the corpus, readable by all of them. The ledger records every composition that ran, what it produced and what it was worth. The first two make the teams possible. The third is what turns them into a business, because a capability nobody counted is a capability nobody improves.
What is true today
—
Five briefs, each naming a job the house can be paid for. Membership is computed: the seat goes to the agent most specialised in that faculty, and no agent takes two seats. Where the register holds nobody, the seat says so rather than being quietly filled.
Every agent in the estate, what it was written to do, and which tables it stands on.
What the house holds, what it would cost to produce again, and what it is worth by five methods that share no assumptions.
Priced on authored, distinct source only. The component library that ships with every generated app — six hundred and fifty-eight thousand lines of it across the fleet — is excluded, as is every file byte-identical to one another application already carries. Two routes, and they are never averaged.
The catalogue figure this replaces was every item's shelf price multiplied by exactly one buyer each. Reproduction costs more than that figure, which is the tell: when producing something again costs several times its stated catalogue value, the catalogue is mispriced, not the asset.
—
The shared memory. One voice, and the reason a composition here cannot be assembled anywhere else.
A hundred books on emergence and developmental facilitation, with a hundred and one narrations recorded. They are held — two thousand and seventy-nine citation markers are still open, and the house does not ship a book that cites something it has not checked.
They are also the natural first shelf for every agent to stand on, and the reason is the subject matter rather than the count: The Practice and The Reading Room are both composed almost entirely of faculties these hundred books are about. An agent holding a developmental passage, citing the volume that describes that passage, is the single least reproducible thing in this estate.
Resolving two thousand and seventy-nine citations is the cheapest work in this document with the largest multiple attached, because it unlocks a hundred books and the shelf every agent would stand on at the same time.
—