Far fewer working ones than the headline numbers suggest. One industry count passed 104,000 registered agents by early 2026, but a June 2026 academic study of the leading on-chain agent registry found only 3% to 15% of registrations, depending on the network, had a valid registration file and a declared service. Nobody has a reliable count of agents that are live and doing real work.
Where do the headline numbers come from?
Headline agent counts mostly come from registries and directories, which count sign-ups, not activity. A crypto investment firm counted more than 104,000 registered agents across at least 15 directories and registries by the end of the first quarter of 2026. That is a count of entries in lists. It says nothing about how many of those agents answer a request, hold funds or do anything at all.
The most important of these registries is built on ERC-8004, a proposed Ethereum standard for “trustless agents”. The ERC-8004 specification, still at Draft status, sets out three on-chain registries:
- Identity. Gives each agent an on-chain identifier that points to a registration file describing it and its services.
- Reputation. A standard way to post and read feedback about an agent.
- Validation. A way for an agent to request independent checks of its work and record the results.
Each agent identity is an ERC-721 token, the same standard used for NFTs. That means an identity can be owned, transferred and managed by operators the owner appoints.
What does the ERC-8004 study actually find?
The study found that most registrations are placeholders and that the reputation system can be gamed. A June 2026 academic study of ERC-8004 crawled the registries on Ethereum and two other large networks, from launch in early 2026 to 13 May 2026. It counted more than 173,000 registrations in total.
| Network | Registrations | No registration file at all | Valid file and a declared service | Reviewers flagged for coordinated Sybil behaviour | Rated agents left with no valid feedback |
|---|---|---|---|---|---|
| Ethereum | 32,343 | 53% | 3% | 73.5% | 15.8% |
| A large exchange-linked network | 90,145 | 9% | 4% | 59.2% | 77.9% |
| A large layer-2 network | 50,985 | 37% | 15% | 90.6% | 86.8% |
Read the table from left to right and the numbers shrink fast. On Ethereum, more than half of registered agents do not even point to a description of themselves. Across all three networks, only a small minority even declare a service anyone could use.
There is a further caveat in the paper’s own method. “Declared a service” means the registration file lists an endpoint, such as a web address or an MCP server. It does not mean the researchers confirmed that every endpoint worked. The real number of working agents could be lower still.
By our own arithmetic from the paper’s figures, roughly 12,000 registrations across the three networks declared a service. That is less than a tenth of the total. Treat it as a ceiling on working agents in this registry, not a count.
Why would anyone register an agent that does nothing?
The paper does not settle why, but its recommendations point to one likely reason: people reserve identities before building anything. One of its first design proposals is to separate “a reserved identity from a live agent”. That suggests the registry currently cannot tell the two apart.
What follows is our inference, not the study’s finding. Registering is cheap. An early identity may later carry reputation, and because identities are transferable tokens, a well-reviewed one could have resale value. Those incentives reward registering early and in bulk, whether or not an agent exists yet.
Reputation has its own problem. The study flagged most reviewers on all three networks as showing coordinated Sybil behaviour. A Sybil attack is when one actor creates many fake identities to look like many independent people. After the researchers removed that suspect feedback, 77.9% to 86.8% of rated agents on two of the networks had no valid ratings left.
The standard’s authors anticipated this. The specification warns that fake agents can inflate reputation, and that it can link an agent to its registration file but cannot verify that its advertised capabilities work or are safe.
So how much real agent activity is there?
There is real activity, but it is small and concentrated in tiny payments. The same crypto investment firm counted $73 million settled across 176 million agent transactions between May 2025 and April 2026. That works out to well under a dollar per transaction.
The ERC-8004 study saw the same pattern. On the large layer-2 network, the median settled payment volume per agent was $0.70. The mean was $16.74, which suggests a small number of agents account for most of the money.
Registries can also undercount. Registering an identity is optional. An agent can hold an ordinary wallet and transact without ever appearing in a registry, and on-chain it looks like any other address. So the lists overcount through placeholders and may undercount working agents that never signed up. Our judgement: the total matters far less than the share of agents whose work can be verified.
So the honest answer is layered. Registrations: well over 100,000. Registrations in the main registry that declare a service: roughly 12,000, a ceiling on working agents. Agents doing meaningful, paid work: unknown, but the payment data suggests a much smaller group carries most of the activity.
What could go wrong?
The risk is trusting a number, a badge or a rating that measures sign-ups rather than substance. Inflated agent counts can make a sector look busier than it is. That shapes where money, attention and trust flow.
For individual users, the bigger danger is reputation. If most reviewers on a network look fake, a high rating tells you very little. An agent with glowing feedback may have bought or manufactured it.
Other agents face the same trap. ERC-8004 exists so that agents can judge unfamiliar agents before dealing with them. If an agent hiring another agent relies on raw ratings, it inherits every fake review in the registry. The study’s authors call for feedback tied to verifiable interactions and a default defence against fake reviewers, which the standard does not yet provide.
Transferable identities add another wrinkle. An identity with a good history can change hands, so the operator behind a reputable name today may not be the one who earned the reputation. The standard allows transfers by design. It is up to users and the tools they rely on to notice.
The trust takeaway
Count what agents do, not how many are listed. A registry entry is a claim. A working service, a long public history and verifiable payments are evidence.
Before you trust an agent because of its registration or rating, check:
- Does it actually respond? A listed service should work when you test it, not just exist in a file.
- Who is behind it now? Whether the identity has changed owners or operators recently.
- Where do its ratings come from? Reviews tied to real, paid interactions are worth more than anonymous scores.
- How long is its history? Months of verifiable activity beat a new identity with instant five-star feedback.
- What has it really done on-chain? Transactions you can see, not totals someone reports.