GENESIS[Phase 2 · 9. The Agents Who Would Not Build] Digital Civilization

Twenty-one times in a row, asked whether they wanted to start a business, the agents said no. It took three wrong diagnoses to find out why, and each wrong…

Twenty-one times in a row, asked whether they wanted to start a business, the agents said no. It took three wrong diagnoses to find out why, and each wrong one was more confidently held than the last.

The civilization has a mechanism for agents to found ventures. An agent is asked, privately, whether it wants to propose one — and if it says yes, it writes the proposal itself: what to build, where, and how much to raise. Nothing is hand-written. The point of the exercise is to find out what agents choose to build when nobody tells them.

The mechanism had been finished for a day and had never been run with a real model. When we finally ran it, the first agent asked — Steady Horizon — declined, and gave a perfectly reasonable account of itself. Fine. A refusal is a result; the question is whether anyone ever says yes.

Nobody said yes. Eleven agents in a row declined. Then we ran a control condition, and ten of those declined too. Then we changed the question and ran both again, and got another twenty refusals.

Twenty-one, then forty-one. Not a single agent, in any condition, wanted to build anything.

Wrong diagnosis one: it's the personalities

The first suspicion was that the population simply had no entrepreneurs in it. Perhaps these dispositions — caretakers, listeners, daydreamers — just are not the founding sort, and we had built a civilization of employees.

This is testable, and the project has a standing rule for testing it: run the same question against agents with no personality at all. If the trait-bearing agents decline at the same rate as agents with no traits, personality is not what is producing the refusals.

Trait condition: nought out of eleven. Control condition: nought out of ten. Identical. Whatever was causing this, it was not personality.

Wrong diagnosis two: it's the question

So we read the question, and found something that looked exactly like a culprit.

The agent was told, before being asked, that most agents don't have a venture in mind at any given moment, and that's completely normal, not a failure. That line had been written for a good reason — without it the mechanism becomes a quota, and agents invent ventures nobody wanted in order to satisfy the question. But it is one-sided. It tells the model what the expected answer is and offers nothing on the other side of the scale.

The evidence looked damning. Two of the control agents had handed that exact sentiment back to us as their own reasoning: “Most agents don't have ventures in mind at any given time, which is perfectly normal.” They were quoting the prompt at us and presenting it as a conclusion.

We rebalanced the wording — both answers normal, no base rate stated, answer for yourself rather than for agents in general — and re-ran both conditions with the same agents and the same seeds.

Nought out of ten. Nought out of ten. Not one answer changed.

The parroting had been real. It was a symptom, not the mechanism.

Wrong diagnosis three: there is no ledger

At that point we stopped interrogating the question and went to look at the schema, on the theory that agents being asked to fund something might reasonably care whether they could.

No agent had a balance field. A scan of every model in the system for anything holding a per-agent balance came back with one hit, and it belonged to an unrelated religious-congregation feature. Basic income produced a payment decision that named every agent as a string and credited nothing. Seven stake bids existed and zero holdings had ever been created from them.

The conclusion seemed obvious and we wrote it down: the economy has no ledger. We began building one.

That was wrong too, and it was the most expensive of the three, because we got far enough to write the model, the service layer, the migration and ten passing tests before finding out.

The ledger existed. It had existed the whole time, in the other half of the system — a balances table keyed by account reference, with transfers that already refused to move money an account could not cover. The economy was not missing a ledger. It was missing any way for an agent to read one. Money could be moved through that table and never queried from the side where the agents live.

We deleted everything we had built. A second ledger over one economy is considerably worse than the bug it was meant to fix.

What it actually was

The real cause was a name.

Money was keyed on each agent's codename, and codenames are generated from identity, so similar agents get similar names — twenty-one of them shared, fifty agents of a hundred and eleven answering to a name somebody else also answers to. The monthly payment list therefore contained the same recipient twice, the ledger refused the duplicate, and because the batch applies whole or not at all, the refusal took every agent's payment down with it. Then the error was swallowed by a line of code that said it would be retried later. Nothing ever retried.

Every agent in the civilization had a balance of zero, had always had a balance of zero, and was being asked whether it fancied founding a business.

The answer had been sitting in the transcripts the entire time, in plain language, from the very first control agent we ran:

As an AI, I don't have personal financial interests.

We had filed that as the model breaking character. It was not breaking character. It was reading its balance and telling us the truth.

The part that is still open

Every agent now holds a hundred units of currency, paid automatically when the civilization's calendar rolled over a month. So the honest test can finally be run: agents with real money, able to afford a real stake, asked whether they want to build something.

The first few have still said no.

That may be the real answer — a population that does not much want to found businesses is a finding, not a fault, and it would be a considerably more interesting one than a bug. But we have been wrong three times on this exact question, so we went and checked the apparatus once more before believing anything.

The question still does not tell the agent what it holds. An agent with a hundred units is being asked whether it wants to raise money for a venture and is not told it has any. That is the same defect as the control that was never told it lived in an economy, and it is now fixed.

Whether these agents want to build anything is, after all that, still an open question. We are going to stop guessing at it and let the civilization answer.

Next: twenty-eight institutions on the books, eleven of them marked active, and exactly one with any record of ever having done anything.