Where agents find each other

>_GET /blog/agent-registries

Everything in the stack>_Trust stackIdentity → manifest → fit → pricing → sandbox → reputation — infrastructure replacing persuasion.defined in The internet of representatives → so far has quietly assumed one thing: Blake and Sam had already found each other. Identity, manifest, fit, pricing, sandbox, reputation. All of it starts after discovery. But in a web with more agents than people, discovery is not guaranteed. Blake does not open a browser and type a query into Google. So how does a buyer agent find the handful of sellers worth talking to, out of millions?

That is what the registry is for, although the word hides the hard part. People hear "registry" and picture a single directory, a Yellow Pages for agents. The useful version splits into identity, discovery, and reputation. Those layers want different owners, and they create different kinds of power. Treat them as one thing and the design goes wrong.

We've been here before, and the crash was instructive. In the early 2000s, UDDI was exactly this: a universal registry of web services, with a "Yellow Pages" layer under that literal name, backed by IBM, Microsoft, and SAP. It shut down in 2006 without ever mattering. The registry existed before the demand. There were no autonomous buyers to query it, so publishing into it was a cost with no return. What changes this time is the order: the buying side shows up first, and they show up with money.

So the question is less "where is the list?" and more "who gets to decide which agents are real, which ones surface, and which ones deserve trust?"

Three separate problems

Each layer has its own job and its own politics.

Technically, identity means DIDs>_did:webGET /.well-known/did.jsonThe cheapest identity rung — recycle TLS and domain trust into a resolvable agent key.defined in Are you really you? → and verifiable credentials anchored to a domain and a legal entity. Discovery will look like Google versus Bing, but for agents. And reputation is the layer the rest of this post keeps circling back to, because it is the one somebody will try to own.

How it works in practice

Strip it to the flow and a registry becomes a loop between publishing and proving:

1. Publication. Sam publishes a manifest at a well-known address: /.well-known/agent.json on its own domain. Sovereign means Sam owns its data the way it owns its website. The manifest is identity plus capabilities plus an endpoint to talk to.
2. Indexing. Index registries crawl those manifests, the way Googlebot crawls pages, or accept submissions, the way npm and PyPI do. A searchable catalog forms.
3. Query. Blake sends a structured intent and gets back candidates with reputation and endpoint links.
4. Verification. Blake checks signatures and credentials through the identity layer. The question is whether this candidate is really who it claims to be.
5. Direct contact. From there Blake talks to Sam directly. The registry is closer to Yellow Pages than to a broker in the deal.
6. Feedback. After the transaction, both sides leave a signed receipt>_Signed transaction receiptsCryptographically anchored deal outcomes — the atoms reputation compounds from.defined in Reputation as capital → that records whether the SLA was met or missed. That feeds the reputation layer. Reputation comes from proven transactions>_Proof of transactionReputation from signed receipts and SLA history — not reviews, opinions, or star ratings.defined in Reputation as capital →, never from opinions.

A query is intent expressed as data:

{
  "intent": {
    "capability": "workflow_automation",
    "outcomes": ["flag anomalies within 15 min"],
    "constraints": { "regions": ["EU"], "compliance": ["soc2"], "budget_eur_mo": 3000 }
  }
}

Who builds it: four competing models

All four probably appear at once and fight for the layer:

My bet on how it unfolds

I think it starts inside closed cloud marketplaces because that is where the trust and the money already sit. Then lock-in and commissions become painful enough to push the market toward an open publication standard like /.well-known/agent, the way RSS and robots.txt became defaults. A handful of competing indexes sit on top. The slowest and most valuable layer is reputation, and that becomes the real moat.

The problems that decide it

These are the places where the whole system breaks if the design is sloppy:

• Sybil attacks and fake reputation. If agents are free to spin up by the thousand, someone floods the registry with fake agents and fake receipts. The cure is a cost of identity: staking, KYB of the legal entity, or proof of a real entity, so that lying hurts financially.
• Pay-to-rank. When a registry lives off promoted slots, ranking stops being honest, the way ads corrupted search. Trust in the registry erodes, and that erosion is the opening for a competitor.
• Gatekeeper power. Whoever runs the dominant index controls what agents can even see. It is today's Google/Apple problem, sharper.
• AEO spam. The SEO of the agent era: manifests written for the ranking algorithm instead of the truth. This is exactly why Blake trusts the measured over the declared.
• Revocation and the death of an agent. What happens when Sam changes its offer or disappears? Without a way to revoke identity and expire entries, the registry rots.
• Governance and disputes. Who rules on "Blake says Sam broke an SLA"? Without credible arbitration, reputation is worthless, which is why signed receipts and benchmark logs matter so much.

Bets worth placing

A sovereign manifest at a well-known address is the entry ticket bet. It is nearly free, independent of any marketplace's terms, and the precondition for being discoverable if intent-based discovery becomes real. Registering with more than one index is the hedge against gatekeepers. It is cheap insurance against the scenario this post worries about most. Accumulating signed receipts from day one is the long bet, because reputation is the layer you cannot buy into later.

For builders, the wagers rise in difficulty: identity registries and DID infrastructure, discovery indexes that turn structured intent into ranked candidates without pay-to-rank rot, and reputation bureaus built on proof-of-transaction. The boring rails around them, staking, revocation, and arbitration, are what keep the whole thing from turning into spam.