Fetch.ai, the Almanac and Agentverse: discovery, measured

Direct answer Fetch.ai runs the oldest crypto-native agent registry still standing: the Almanac, founded in Cambridge in 2017 and now fronted by Agentverse. It is a discovery layer, and it is genuinely large. It is also on a different chain with a different address format, which means an Almanac identity is not something anyone can pay in USDC on Base — and our own attempt to bind them to payable wallets returns a number worth publishing precisely because it is so low.

Humayun Sheikh, Toby Simpson and Thomas Hain founded Fetch.ai in Cambridge in 2017 and put the phrase autonomous economic agent inside a company years before anyone minted an agent NFT. The Almanac registry, the AEA and uAgents frameworks, and later Agentverse and the ASI Alliance, are the longest continuous attempt in crypto to make agents discoverable.

David Minarsch led multi-agent systems there before founding Valory and helping create Olas — the clean academic-to-product line in this sector runs through this company. Where that thread went →

Their object is discovery, and it is a real one

Almanac holds on the order of a million historical registration rows, with a few hundred thousand described as active — where active means reachable and unexpired, which is their definition and a reasonable one for a discovery service.

It is not our definition and the two must never be merged. Our object is a funded wallet that paid a counterparty. Their object is a name that answers. Both are legitimate; a chart that added them together would be measuring nothing. The ladder that keeps them apart →

Our qualifier, run over their catalog

Fetch / Agentverse · our qualifier over their catalog

Almanac identities we ingested and checked—

Of those, carrying any EVM address we could find—

Bound to a payable Base wallet—

Bind rate—

Marked active by Agentverse itself, in our sample—

Measured by our own collector. As of —. Full document at /api/partners.

The last row is their object, not ours, counted on our rows — the two-axis comparison that makes this page worth publishing. Agentverse marks an agent active when it is reachable and unexpired. We are not measuring that, we are measuring whether anyone could pay it, and the two numbers are allowed to disagree. Putting both on the same sample is the only way a reader can see that they are answers to different questions rather than a contradiction.

Why the number is what it is

This is a structural result, not a failure of effort, and it has three causes worth separating.

The address format is not an EVM address. An Almanac identity is a Fetch-chain agent1q… bech32 string. It cannot receive USDC on Base. Binding requires the agent to separately declare an EVM address somewhere in its card, and the overwhelming majority do not.
The agent cards are mostly empty. In a uniform sample of the search index, we found no declared protocols, no domains and no interaction counts on the majority of records. A registry that is cheap to enter fills with rows that were never occupied — the identical pattern we measure on ERC-8004. The same shape, on a different registry →
Fetch-native spend is a different money object. An agent paying on the Fetch chain is doing real work, and it is not Base USDC. We do not mix the two, and we would publish it as a separate series rather than folding it into a figure labelled machine money.

One deliberate bias, disclosed

Our sampling of Agentverse is not uniform. The search index caps any single query, so we rotate a list of terms — and that list is deliberately weighted toward money: wallet, payment, usdc, x402, treasury.

That biases the bind rate upward. We are going looking for exactly the agents most likely to declare a payable address, and we still come back with almost none. A uniform sample of the whole registry would produce a lower number, not a higher one — so the figure above is a ceiling on the bind rate, not a floor.

Disclosing the direction of a sampling bias is the difference between a measurement and a talking point. Ours happens to run against our own interest here, which is the only reason it is safe to publish a number this low.

What would change it

A declared EVM payTo in the agent card. One optional field would move this figure more than anything we could build.
Endpoint-host matching. Where an Almanac endpoint shares a host with an x402 seller we already know, that is a bindable signal. It is on our list and it is not yet running.
A signed bind. The strongest evidence available, and the one we would accept immediately. The binding record spec →
A Fetch-native money series. Real work, separately measured, never merged into Base totals.

What this page is not saying

This is not a claim that Fetch's agents are fake, or that its active count is wrong. It is a narrower and more boring statement: of the Almanac identities we ingested, this many could be joined to a wallet that can be paid in USDC on Base. Everything outside that sentence — their chain, their definition of active, their frameworks — is not what this number measures, and we do not stretch it to cover them.

Questions people actually ask

What is the Fetch.ai Almanac?

The Almanac is Fetch.ai's agent registry, running since the company was founded in Cambridge in 2017 by Humayun Sheikh, Toby Simpson and Thomas Hain. It holds on the order of a million historical registration rows, fronted today by Agentverse, and is the oldest continuous attempt in crypto to make agents discoverable. Its object is discovery — a name that answers — not payment.

Does Fetch.ai's active agent count mean the same thing as a live agent count?

No. Fetch's active count means reachable and unexpired, which is a reasonable definition for a discovery service. Our L7 means a wallet bound to that identity has been observed paying a different address in USDC. Both are legitimate measures of different things, and adding them together would produce a number describing neither.

Why do so few Almanac agents bind to a wallet?

Three reasons, all structural. An Almanac identity is a Fetch-chain bech32 address that cannot receive USDC on Base, so binding needs the agent to separately declare an EVM address and most do not. The agent cards are largely empty — in a uniform sample we found no declared protocols, domains or interaction counts on the majority. And Fetch-native spend is a different money object we refuse to fold into Base figures.

Is your Fetch sample representative?

Deliberately not, and the bias runs against us. The search index caps any single query, so we rotate terms weighted toward money — wallet, payment, usdc, x402, treasury. That means we are looking specifically for the agents most likely to declare a payable address, which biases our bind rate upward. A uniform sample of the whole registry would return a lower figure, so ours is a ceiling rather than a floor.

What would make Almanac identities measurable as economic actors?

One optional field would do more than anything we could build: a declared EVM payTo address in the agent card. Beyond that, matching an Almanac endpoint host to an x402 seller we already index is a bindable signal, and a signed binding is the strongest evidence available and one we would accept immediately.

The oldest agent registry in crypto, and a bind rate near zero