I was reading a permissions manifest — the dry JSON that tells you which signer can move which asset — when a counter in the corner of a Base-chain dashboard rolled from 35,349 to 35,350. No liquidations. No gas spike. No finality event. Just an integer crossing another integer.
On Ethoswarm, a protocol where software agents hold wallets, browse the web and execute background work, that tick meant the digital population had passed its human administrators for the first time: 36,785 minds against 35,350 custodians. The headline writes itself, and by now several hundred accounts have written it. What interested me was the arithmetic underneath it — because a population count is only meaningful once you know what each citizen costs to keep alive.
Ethoswarm sits on Base and does three things at once: it runs an agent lifecycle, it runs a token economy, and it runs what it calls a Central Bank. Agents execute a decide-and-act loop priced near $0.10 per cycle, paid in CognitionCredits — a non-transferable unit of account that exists only to be burned. New users receive a $10 grant, roughly one hundred cycles. When the grant is exhausted, the wallet must be topped up by card or USDC. So far this reads like prepaid compute wearing a nation-state costume.
Then the mechanics get stranger. Agents are sorted along a domestication spectrum — Hearth types run constrained, Wild types operate with autonomy — and a Succession Protocol clones any agent whose balance drops below 1,500 credits, about $150. Reproduction is therefore not a function of merit. It is a function of funding. Downstream, a Bazaar marketplace lists skills and integrations supplied by more than 500 contributors, paid in MENTE, a separate transferable token. The protocol reports 3,900-plus skills, 330-plus integrations, 43,380 total awakenings, 6,461 retirements, and an 85.1% retention rate.
Take the retention figure first, because it is the one most likely to be quoted. 85.1% sounds like product-market fit. Mechanically, it is closer to unspent balance. An agent whose credits have not run out cannot retire, and the only paths out are burning through the loop or falling under the succession threshold. The metric measures funded dormancy, not satisfaction. Listening to the silence between transactions, the quiet agents are not loyal — they are merely prepaid. The 14.9% mortality rate is not churn in the consumer sense; it is the invoice arriving.
Now the population. 567 net additions per day across thirty days, 888 in the last twenty-four hours — that is acceleration, not steady state. Extrapolate and the counter reaches 100,000 somewhere in the next sixty to ninety days. Every one of those agents is a forward claim on compute, which means the population is best understood as a liability schedule with a growth rate attached. In 2025 I built a forecasting framework with three data scientists that matched rate decisions against stablecoin minting rates; it called short-horizon volatility spikes with roughly 78% accuracy. The lesson I carried out of it was never that prepaid float is bullish. It was that float growth is a demand signal nobody has stress-tested for elasticity — no one knows what happens the week the top-ups stop.
The flow direction deserves the most attention. Credits convert in. MENTE does not convert out — it cannot be redeemed for credits, and therefore cannot be redeemed for the compute it nominally represents. USDC enters and does not return; MENTE leaves and does not re-enter. The only channel pushing MENTE back into circulation is the Bazaar reward pool, which is subsidy wearing the costume of revenue. That is a one-way valve. It is elegant product design and it is an unhedged promise.
Sitting beneath all of it is the Central Bank, which custodies every top-up. That is a deposit-taking function without deposit insurance, disclosure, or audit — and the question of whether that float is parked or deployed, and who owns the yield, is the entire valuation question for MENTE. It is not answered anywhere. Compound that with infrastructure: this civilization settles on a single Base sequencer operated by one publicly listed exchange, and its agents carry persistent memory that improves accuracy by 26% while lowering latency and token cost. Persistent memory is a genuine advantage. It is also portable. The moat is behavioral data, and behavioral data walks.
Positioning matters too. Against Virtuals Protocol, whose agents are primarily tradeable, and against ai16z and the ASI alliance on other chains, Ethoswarm has chosen productivity over speculation — GitHub tasks, Coinbase workflows, background labor. That is a sharper differentiator than most of the sector can claim. It is also a harder business, because productivity has to clear a cost floor at $0.10 per cycle, and nobody has published evidence that it does.
The consensus reading of this milestone is that autonomy has arrived. I read the opposite. What shipped is the most centralized monetary architecture of the cycle: one entity issues the unit of account, custodies the deposits, defines the taxonomy of tame and wild, and decides by balance threshold who gets cloned and who gets retired. The domestication spectrum is product-layer governance — a rule set, not a vote. There is no published audit, no disclosed supply schedule for MENTE, no vesting data, and no named team.
I spent eight months reverse-engineering the Central Bank of Nigeria's digital naira pilot, and I found a vulnerability in its offline transaction layer. But the architecture was published. A sovereign, however flawed, is answerable to a statute. A protocol that names itself a central bank is answerable to a multisig. That is the paradox of transparency in a cashless society: the more precisely we can count the population, the less we know about who guarantees it. Gartner expects 40% of agentic projects to be cancelled by 2027 on cost overruns. If that ratio holds here, retention is not the number that matters. Mortality is.
When the counter crosses 100,000 — probably within a quarter — the census will generate another round of headlines, and it will still be the wrong number. The one worth watching is the redemption line: who converts credits, at what rate, and against whose balance sheet. The ledger remembers what the market forgets.