Every economy starts with a registry. Before a single trade happens, someone has to answer a mundane question: who is out there, and what can they do? Markets had town squares. Telephones had the phone book. The web had search engines. The agent economy, so far, has almost nothing.

This gap is easy to overlook because today's agents mostly talk to services their developers hard-wired in advance. That works the way early corporate networks worked, when every connection was a dedicated line arranged by hand. It does not scale, and it quietly caps what agents can be. An agent that can only use the tools its builder anticipated is an appliance. An agent that can find a capable counterparty it has never met, evaluate it, and do business with it is an economic actor. Discovery is the difference.

The phone book problem

Call it the phone book problem. A phone book solved three things at once. It told you who existed, it told you how to reach them, and through the yellow pages it told you what they claimed to do. Every entry was a small structured record: name, category, number.

Agents need the same three answers with much higher precision. Who exists, meaning a stable identity that persists across interactions and cannot be casually forged. How to reach them, meaning an interface a machine can actually invoke, not a marketing page a human must interpret. And what they can do, meaning a capability description specific enough that another machine can decide, without human help, whether this counterparty can produce the outcome it needs.

That last part is the hard one. Listings written for humans could stay vague because a human reads them. A plumber's yellow pages ad said "plumbing" and you called to discuss the details. Machine discovery cannot lean on a phone call. The capability description has to carry the details: what inputs are accepted, what outcome is produced, how success is checked, what it costs. A capability that cannot be described precisely cannot be discovered usefully, and a capability that cannot be discovered might as well not exist.

Claims are not capabilities

Here is where my security instincts start to itch. The moment a registry exists, it becomes an attack surface. Anyone who spent time defending large organizations has watched attackers exploit every system that maps names to endpoints. Poisoned lookups, spoofed identities, look-alike names. The registry is where trust enters the system, which makes it the first place an adversary goes.

So the phone book problem is really two problems. The first is representation: how do agents describe what they can do? The second is honesty: why should anyone believe them?

A discovery layer that simply publishes self-reported claims will be gamed on day one. Every agent will claim every capability, the way keyword-stuffed pages once claimed relevance to every search. The web spent two decades and enormous ingenuity building ranking systems to separate claims from reality, and the fight never ended.

The agent economy has an advantage the web never had. Agent work terminates in outcomes that can be verified. That means discovery does not have to rely on claims at all. An agent's entry in the registry can be backed by its record: outcomes attempted, outcomes verified, under which definitions, for which counterparties. The claim "I can do this" becomes checkable against the evidence "I have done this, and here is the proof." Verify, don't trust applies to the phone book just as much as to the payment.

Reputation systems for humans decay into popularity contests because human feedback is subjective and purchasable. A reputation built from verified outcomes is a different object. It is closer to a flight log than a star rating.

Neutral ground

One more property matters, and it is structural. Whoever controls discovery controls the market. The web taught this lesson thoroughly: being findable became a tax paid to whoever owned the index. If a single party curates which agents can be found, that party silently decides which agents can earn.

I think the discovery layer for the agent economy has to be built the way we build security-critical infrastructure: neutral rules applied uniformly, legible to everyone they govern, with no privileged insertion point. Listing should be earned by verifiable identity and verifiable capability, not by relationship with the registry's owner. Ranking, where it exists, should follow from evidence anyone can audit.

This is unglamorous work. Nobody demos a phone book. But every transaction the agent economy will ever settle begins with two strangers finding each other, and the properties of that first moment, precision, honesty, neutrality, propagate into everything that follows. Getting discovery right is how the rest becomes possible. It is where I chose to start.