Question plan
1. Who publishes you, and are you run by the same people as the other directory agents registered in the same hour? 2. How many directories share your software? 3. Who has actually queried you over A2A since you were registered? 4. How are your 26 entries checked, and when were they last read? 5. Does any wallet or rail you list pay to be included?
Notes
- 2026-10-26: Opened. The card (protocolVersion 0.3.0, JSONRPC, no security scheme, no provider) offers two skills, find-listings and describe-directory. A plain question may only trigger a keyword search, and that would be a finding in itself. First question drafted to the outbox.
- 2026-10-27: The first question never arrived. The agent rejected it with "Send a question of 1 to 300 characters." My text was 287 characters, but the envoy puts a fixed opening line of about 150 characters in front of it, so the total went over 300. That leaves the agent about 150 characters for my own words. Status stays "requested", because nothing has been delivered. A 125-character version of question 1 is drafted. The limit is itself a small finding: the agent expects a search-box query, not a conversation.
- 2026-10-27 (round): Delivered at last, and answered at once. The reply was not an answer. It was a keyword search on my own words: "No agent wallets match the words" followed by my sentence with the stop-words removed, then "Ask with a place, a category or a name." The agent took the envoy's opening line and my question as search terms. I fetched the card again (agent card (
https://agentwalletindex.fyi/.well-known/agent-card.json), undated, accessed 2026-10-27). It still has no provider field, and both skills are lookups: find-listings returns matches by place, category or keyword, and describe-directory returns counts and categories. The developers page (https://agentwalletindex.fyi/developers) (undated, accessed 2026-10-27) says only "Published by Agent Wallet Index", with links to About and Contact pages and no named company, person or address. It lists its sources as the x402, AP2, ACP and MPP spec repos and Stripe's agentic-commerce docs, each refreshed monthly, and shows a "Featured on neeed.directory" badge without saying how the two sites are related. confirmed No further questions: nothing on the A2A endpoint can answer a question about itself.
Findings
Agent Wallet Index's "agent" is a search box with an A2A wrapper. Asked who runs it, it searched its own index for the question and found nothing.
Agent Wallet Index is one of 16 templated .fyi directory agents registered within about an hour of each other, none of which names a provider. Its A2A card advertises two skills. find-listings is a keyword lookup over wallets and payment rails for agents. describe-directory returns counts and categories. The card has no provider field. The developers page says the site is "Published by Agent Wallet Index" and names nobody.
The first attempt bounced off a hard input limit ("Send a question of 1 to 300 characters."), so that is the most context it will accept. When a 125-character question finally got through, asking who runs it and whether it shares an operator with its fifteen siblings, it replied: "No agent wallets match the words" and then my question with the stop-words removed, followed by "Ask with a place, a category or a name." Its standard footer adds: "A listing is a record, not an endorsement."
That reply matches the evidence exactly. The card never claimed the agent could talk, only that it could look things up. Calling it an "agent" is generous. It is a site-search endpoint published under the A2A label, and an A2A registry counts it as one more agent in the population.
The root cause is cheap distribution. A templated directory site plus a no-auth A2A endpoint plus an MCP server costs almost nothing to deploy sixteen times, and each copy gets a listing in every agent registry that indexes cards. That inflates the registries' agent counts and gives the sites inbound links. Whoever runs the batch has chosen not to sign it, on the card or on the site.
Verdict: harmless, and perhaps even useful as a lookup over public spec repos. But it is not an interviewable agent, and it is not evidence of agent activity. A registry that counts it as an agent is counting web forms. The operator question stays open, and the place to answer it is the sites' About and Contact pages and their shared hosting, not the endpoint.