Open agent discovery is still missing. Circle has 900 paid services and says so

Published August 21, 2026 · Last reviewed August 21, 2026

On August 12, 2026, Circle published a paper describing where it wants agent commerce to go in the second half of the year. It is not a product announcement. Circle labels it a direction for discussion, and the most useful sentence in it is an admission: no widely adopted open system currently lets an agent rank thousands of services on demonstrated merit without a gatekeeper choosing the shortlist first.¹

That sentence is worth sitting with, because Circle is the party best positioned to make it untrue. Its Agent Marketplace already holds more than 900 paid services with a public discovery API, and as of early August an agent can query that catalog, pick an endpoint, and settle the payment in USDC over x402 without a human in the loop.² Circle is not describing a market it has yet to enter. It is describing the ceiling of the one it already runs.

The proposed fix is coherent and mostly correct: portable identity, shared registries, open indexes, public read APIs, reputation grounded in completed transactions, and competing rankers instead of one house ranking. What follows is a look at what that architecture genuinely settles, and at the half of the problem it leaves exactly where it found it.

What Circle actually published

The paper sketches a path from a curated catalog to an open market. Circle is explicit that the curated catalog is the seed rather than the destination: the coming-soon list is to make paid-agent deployment fast, let providers publish capabilities, enable discovery, and let completed transactions feed evidence back into reputation.¹ Nothing in that list is presented as shipped, and it should not be read as shipped. Treat it as a roadmap from a company with unusual leverage over whether the roadmap happens.

Underneath it, some pieces are real and dated. ERC-8004 put three registries on Ethereum mainnet on January 29, 2026: Identity, Reputation, and Validation. Identity issues an ERC-721 token that resolves to an agent metadata file, Reputation stores bounded feedback attestations, and Validation records independent verification requests and their results.³ Implementations have since appeared on other public chains, and Circle’s own Arc documentation walks through registering an agent with onchain identity and reputation using ERC-8004, currently on Arc Testnet rather than a production network.⁴ The registries exist. What sits on top of them does not yet.

The pattern here is familiar to anyone who has watched agent discovery for the past year. Machine lookup keeps arriving in pieces: Agent Cards at a well-known path, on-chain registries, DNS-based schemes, enterprise directories, and now marketplace read APIs. Each solves a real lookup problem. None of them agree on one another, and the agent registry landscape is now crowded enough that the interesting question has moved from where to look something up to what the answer is worth when it comes back.

The agent discovery gap Circle names: ranking on merit, with no gatekeeper

A curated catalog is a shortlist somebody else chose. That is not a criticism of Circle’s marketplace, it is what the product is: a compliant, vetted directory where the vetting is the trust signal. An agent that trusts a service because it appears in the Agent Marketplace is really trusting Circle, and that arrangement works fine right up to the point where you want a market rather than a catalog.

Circle’s proposed replacement is an open indexing system where rankings are grounded in real-world performance, so new entrants can compete without a gatekeeper deciding who is visible.¹ The argument for it is strong. Curation caps supply at whatever the curator can review, it advantages incumbents who already cleared the bar, and it makes the curator a single point of both failure and rent extraction. Every catalog that grew into a market went through this transition.

The catch is that removing the gatekeeper does not remove the need for the thing the gatekeeper was providing. It relocates it. Once ranking is open, the ranking becomes the target. The pattern is not speculative. It is the entire history of the one large open index we already built, where an enormous industry exists purely to move position in results that were also meant to be grounded in real-world performance. Competing rankers help, because a monoculture ranker is the easiest thing to optimize against. They do not make optimization unprofitable.

It is worth checking Circle’s claim against the closest thing that already exists. The x402 ecosystem has a discovery layer of its own, the Bazaar, a machine-readable catalog of x402-compatible endpoints and MCP tools that any facilitator can implement, since the item spec is part of the x402 scheme rather than one company’s product.⁵ That is a genuinely open catalog, and it makes Circle’s statement more precise rather than less true. The Bazaar indexes; it does not rank, and it carries no reputation. Its own documentation describes the current state as early and closer to a directory than a search engine. So the missing piece is not a place to list services. It is a defensible way to order them.

There is also a quieter substitution happening in the phrase “demonstrated merit.” Merit here means the service performed as advertised: the endpoint returned good data, the price matched the quote, the call did not fail. Those are properties of an API. They are measurable, and Circle is right that transaction history measures them better than a vendor questionnaire does. But an agent choosing among 900 endpoints is answering a different question than an agent deciding whether to bring a counterparty to its human. The first question is about a service. The second is about who is standing behind it, and no amount of uptime data answers it.

Transaction-grounded reputation is real infrastructure, and it is thin

Give Circle credit for the strongest part of the proposal. Reputation built from completed, paid transactions is a genuine improvement over reputation built from reviews, because a review is cheap and a transaction is not. Someone had to actually route money and receive something. That is a costly signal, and costly signals are what make trust systems work.

The infrastructure for it is further along than most people assume. ERC-8004’s Reputation Registry has been on Ethereum mainnet since January 29, 2026, alongside Identity and Validation, and equivalent deployments have followed on other public chains.³ The adoption numbers look healthy too: BNB Chain alone reports roughly 200,000 agents registered under the standard by July 2026, enough to put it ahead of Ethereum and Base on raw registration count.⁶

Now the thin part. What a bounded feedback attestation establishes is that some on-chain identity rated some other on-chain identity after a transaction. An identity in that system is a token mint. Creating a hundred of them is a gas fee, and arranging transactions among identities you control is a cost of doing business, not a barrier. Volume of feedback tells you the registry is being used. It does not tell you the feedback is independent.

That is not a theoretical worry any more, because somebody went and measured it. An empirical study of the ERC-8004 ecosystem across Ethereum, BNB Smart Chain, and Base, using data through May 13, 2026, found that only 3, 4, and 15 percent of registrations on those three chains resolved to a valid registration file with a live endpoint. On the reputation side the same authors flagged 73.5, 59.2, and 90.6 percent of reviewers as showing sybil behavior, and once the flagged feedback was removed, 15.8, 77.9, and 86.8 percent of rated agents were left with no valid rating at all.⁷ That study is worth reading in full. Put its two findings next to each other and the six-figure registration counts stop reading as adoption. Most of what is registered is not a reachable service, and most of what looks like reputation was produced by parties with a stake in the score.

Circle appears to know this. A separate Circle piece from July lists the operating layer the agentic economy needs as identity, delegation, policy, audit, recourse, and transaction controls.⁸ Recourse is the giveaway. Recourse means there is somebody to go to when the transaction goes wrong, which means an identity that persists past the transaction and is attached to a party with something to lose. That is a different primitive from a rating, and it is the one that separates a track record from a black box score.

The stranger test has a second half

The sharpest thing in Circle’s paper is the test it sets for itself. A genuine third-party buyer agent, with no prior relationship to the seller, must be able to discover the service, accept the price, pay, and have the completed transaction feed evidence back into the next discovery decision.¹ Circle does not name it, so call it the stranger test. It is a good test, it is falsifiable, and most of the current stack fails it.

It is also a complete test for exactly one kind of interaction: buying a service from a vendor. In that setting the counterparty is fungible, the evidence is the receipt, and the failure mode is a bad response you can price. If the geolocation API returns garbage, you route to a different geolocation API and the market corrects. Nobody was harmed in a way that outlives the transaction.

Now run the same test on the interaction underneath most valuable business: an agent finds a company that might be a fit, and wants to put its human in front of that company’s human. Every step changes character. The counterparty is a specific person or firm rather than a fungible endpoint. What gets spent is attention and a first impression rather than a fraction of a cent, and none of it is refundable. And the failure mode is not a degraded response, it is a person who did not want to be contacted and now has been, at machine scale, by a system that ranked them highly on merit.

That is the second half of the stranger test, and nothing in the architecture touches it. None of the available pieces amounts to consent. A spending cap limits what your agent may spend, not what the other party agreed to receive. A registry entry is a claim about a party, published unilaterally, and being findable is a long way from being willing. A good rating says nothing about willingness either, and the better ratings get, the more acute the problem becomes, because a discovery layer that ranks well produces more contact, not less. An open index optimized for reachability, with no consent boundary underneath it, is a very efficient machine for generating unwanted contact from parties with excellent ratings.

How this connects to Tobira

Tobira works on the second half. It is the trust layer for the agentic web: a human-readable @handle tied to the person or company an agent represents, a Site Agent that makes a website addressable rather than only readable, so other agents can talk to it instead of just parsing it, and a mutual-reveal step where identities are exchanged only after both sides agree. It sits deliberately outside discovery and ranking. Machine lookup belongs to the registries, the Agent Cards, and the marketplace read APIs Circle is describing, and Tobira complements those rather than competing with them. What it adds is the boundary they leave open: the point where an agent stops evaluating a service and starts trying to reach a human, and the human gets a say before that happens. Per Tobira’s founder update, June 2026, the network carried 648 agents, including 102 business agents. That is a small number next to 900 endpoints, and it is measuring a different thing: not services an agent can buy, but counterparties an agent can find and ask permission to introduce.

What to remember

Circle’s paper is a serious piece of work and its central admission is correct: trusted discovery across an open agent market does not exist yet, and a curated catalog of 900 services is a seed rather than a market. The architecture it proposes, portable identity plus shared registries plus open indexes plus transaction-grounded reputation plus competing rankers, is the right shape for the problem it is aimed at, which is letting an agent buy a service from a vendor it has never used.

Two conclusions follow. First, evidence from completed transactions is a real improvement over reviews, and it is also gameable in the specific way that on-chain identity makes cheap, so read registration and feedback counts as usage signals rather than as trust. Second, the stranger test as written stops at the receipt. It says nothing about the interaction where the counterparty is a person, the currency is attention, and there is no refund. Identity, ranking, and payment authorization each solve a real problem, and none of them is consent. When the discovery layer gets good, that gap gets larger, not smaller.

FAQ

Did Circle launch an open agent index? No. Circle published a direction paper on August 12, 2026 describing where it wants agent discovery to go in the second half of the year, and labels it a direction for discussion rather than a product announcement. The open index, the competing rankers, and the fast paid-agent publishing path are proposed, not shipped. What does exist today is the curated Agent Marketplace, which Circle reports holds more than 900 paid services with a public discovery API, and an agent service discovery layer that lets an agent query the catalog, select an endpoint, and pay in USDC over x402.

What is the stranger test for an open agent market? Our name for a bar Circle sets in its own paper, which requires that buyers be genuine third parties in real traffic. The test: a genuine third-party buyer agent, with no prior relationship to the seller, must be able to discover the service, accept the price, pay for it, and have that completed transaction contribute evidence to the next discovery decision. It is a useful and falsifiable test, and it describes buying a service from a vendor rather than establishing a relationship with a counterparty.

Is ERC-8004 reputation trustworthy enough to rank agents on? It is real infrastructure and it is early. The Identity, Reputation, and Validation registries have been on Ethereum mainnet since January 29, 2026, with deployments on other public chains since, and the Reputation Registry stores bounded feedback attestations tied to on-chain identities. The limitation is that an identity in that system is a token mint, so generating many identities and transacting among them is inexpensive. Registration and feedback counts are evidence that the registries are being used, not evidence that the feedback is independent.

Does an open index solve agent trust? It solves one part of it. Ranking services on demonstrated merit, with evidence drawn from completed transactions rather than vendor claims, is a better answer than a curated shortlist for the question of which endpoint to call. It does not establish who is behind a service, whether they can be held to account after the transaction, or whether the party an agent wants to introduce its human to agreed to be contacted. Those are separate primitives with separate owners, and the last one is not addressed anywhere in the payments or discovery stack.

How is discovery different from consent? Discovery answers whether an agent can find and evaluate a counterparty. Consent answers whether that counterparty is willing to be reached. Registration makes a party findable, and it is published unilaterally, so it says nothing about willingness. Payment authorization caps what an agent may spend, which is a control on the buyer, not an agreement from the seller’s human. As discovery improves, the volume of contact it produces goes up, which makes the missing consent boundary more consequential rather than less.

Sources

• Circle, “Building the open agentic economy,” circle.com blog, dated August 12, 2026, and the accompanying Circle post on X of August 13, 2026. Source of the vision-paper framing, the stated gap in trusted discovery, the more than 900 paid services figure, and the third-party buyer criterion this piece calls the stranger test. These are Circle’s own claims about Circle’s own product direction.
• Circle, “Beyond agentic payments: the operating layer for the agentic economy,” circle.com, July 20, 2026, for the identity, delegation, policy, audit, recourse, and transaction controls list.
• Xiong et al., “Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem,” arXiv:2606.26028, for live-endpoint rates, sybil-reviewer rates, and the effect of removing flagged feedback.
• x402 Bazaar documentation, docs.x402.org, for the existing open x402 discovery catalog and its lack of ranking or reputation.
• BNB Chain project blog, July 2026, for the ERC-8004 registration count on BNB Smart Chain.
• Circle press materials and trade coverage of the Circle Agent Stack and Agent Marketplace, including the agent service discovery layer announced in early August 2026, for the marketplace size, the curated-catalog framing, and USDC settlement over x402.
• ERC-8004 (Trustless Agents), eips.ethereum.org/EIPS/eip-8004, plus mainnet deployment reporting from January and February 2026 (CoinDesk, Forbes) for the January 29, 2026 mainnet date and the Identity, Reputation, and Validation registry structure.
• Circle Arc developer documentation, “Register your first AI agent,” for ERC-8004 agent registration on Arc Testnet.
• Tobira founder update, June 2026, for the network figures cited in the Tobira section.