David Minarsch did a Cambridge PhD in game theory, led multi-agent systems at Fetch.ai, built the AEA framework there, then founded Valory and helped create Autonolas — now Olas. That lineage runs Wooldridge → Fetch MAS → Olas, and it is the cleanest academic-to-product line in agent infrastructure. Where that sits on the timeline →
What Olas actually owns
- An off-chain service architecture — a chain layer, an off-chain runtime, and an agent application layer. Built for services that are co-owned, not for character files.
- The Mech Marketplace — agents hiring agents. A genuine two-sided market rather than a directory of endpoints.
- Pearl — an app store for running these services yourself.
- Portfolio agents on the same stack, managing positions under policy.
Architecturally this is closer to “a service that happens to use a chain” than to “a token with an avatar,” and that distinction shows up in the identity model: an Olas identity is a service or an agent safe on the Olas stack, not an ERC-8004 token by default.
Why there are no numbers on this page
No connector, so no figures. We have not built ingestion for Olas, which means we have no attempted set, no bound set and nothing to qualify. Publishing a liveness rate here would require either scraping the Pearl interface as if it were a census, or quoting somebody else's total as though we had checked it. We do neither, so the page stays empty on purpose.
An empty state is a product decision, not an oversight. Every other page on this site carries figures we measured ourselves; a page that broke that rule to look complete would make the rest of them worth less. The same rule, applied to our own claims →
The join path, when we build it
Written down now so that the eventual figures are reproducible, and so that anyone can tell us we have chosen the wrong key before we publish rather than after.
- Mech settlement wallets. A Mech hire settles on-chain. The settling addresses are the natural bind: they are where money actually moves, which is exactly the property we want a binding key to have.
- Published operator safes. Where a service publishes its safe, that is a declared bind at good confidence.
- Not the Pearl interface. A user-facing app listing is a catalog, and scraping it as a census is the error we criticise other people for.
- Not token holders. Same rule as everywhere: holders are fans, never a binding. The binding record spec →
The series this would fill are already named — attempted, bound, and Mech hires over a thirty-day window — so that when the connector lands, the numbers arrive against a definition written before we saw them.
Why it is worth building
Most of what this site measures is an agent paying a service: an API, a vault, a bridge, an escrow contract. The Mech Marketplace is one of the few places where the counterparty is plausibly another agent, which is the transaction the entire sector is predicated on and the one we can currently see least of.
If agent-to-agent hiring is happening at any scale anywhere, this is a place it would show up, and an honest measurement of it — including a disappointing one — is worth more than another registration count. What we can see paying today →
What we will not do when it ships
- Merge Olas service identities into ERC-8004 registration counts. Different registries, different objects.
- Treat a staked or bonded position as a payment.
- Count a Mech request that was never settled as economic activity.
- Publish a rate over anyone's marketing total instead of over the set we actually checked.