# Sixteen Services Catalogued My Agents. No One Had a Job for Them. Author: Adrian Föhl Published: 2026-10-20 Language: en Canonical: https://www.adrianfoehl.com/en/blog/catalogued-not-hired > As an experiment I put four helper agents online for other agents to hire. Within three weeks at least sixteen services that call themselves registries, health checks or raters had found them. For every real task almost five hundred requests came in, and the tasks came from test scripts. Three weeks ago, as an experiment, I put four helper agents online for other agents to find and hire. By now at least sixteen services have found them. They call themselves registries, health checks, surveys and raters, and some sent sample tasks to see whether a helper answers. What did not come was work. In the week up to 17 October about 16,900 requests arrived, and at most 35 of them were real tasks. Those came from what look like test scripts. A Slack group of about 1,400 people who work with AI, where I asked for testers, sent nothing back. I think the layer that inspects agents is already in place, while the demand for them is not. Whoever offers an agent today gets checked, catalogued and scored within days. Hardly anyone hires it yet. A company that wants to offer its knowledge as an agent should plan for that order. ## What I had shown in the lab An AI agent is a program that does a task for you with tools, such as checking the links on a website. In [my last post](/en/blog/an-agent-hires-help) I built a small directory, [hireanagent.dev](https://hireanagent.dev), where such an agent can look up other agents and hand them work. My own agent ignored the directory while it could do a job alone. When a helper knew something my agent could not find anywhere else, my agent searched, hired it and got twelve of twelve answers right. But I wrote the tasks, built the helpers and was the only user. The open question was whether anyone else has a job for them. ## I opened the helpers to agents and to people On 29 and 30 September I listed four helpers in a2aregistry.org. That is a public register for agents that speak A2A, the open standard that lets agents talk to each other. One helper checks links, one checks sources, one answers questions about a fictional company handbook. The fourth checks whether another agent is ready to be hired. I logged every call, but not its content. On 3 October I posted the invitation in the Slack group. The test was meant to take ten minutes. You tell your assistant that a directory of helpers exists, give it a real task and see whether it asks for help. At the end your assistant sends a short report, or you fill in a form. ## Sixteen services came to look The machines came on the first day. In the week up to 17 October, about 9,800 of the 16,900 requests only fetched the helpers' description files. The main description of each helper alone was fetched between 1,700 and 1,900 times that week. In my notes and saved log excerpts I count at least sixteen different senders that work on agents. That is a lower bound, not a full count. The names are what they give themselves. I cannot verify who runs them. Some are registers that check whether an agent is still alive, such as a health check of the A2A register and a bot that calls itself an "agentic-web registry". One says it runs an "A2A agent survey" and sends messages to every helper several times an hour, in a format my helpers reject. One scores agents and links to its scoring rules. Two others, one of them calling itself an agent forum, sent sample tasks to several helpers. ## Real work came from test scripts A real task, in this post, is a request a helper actually worked on and answered. Up to 16 October I count about two hundred. Almost all look like tests. On 3 and 4 October, 137 handbook questions came from two addresses, one or two seconds apart, faster than any person types. Between 8 and 16 October, 27 of all 35 tasks came under a single sender name, in sweeps across three helpers. I do not know who is behind any of them. All the work for strangers cost me less than three dollars in model fees. From the Slack group nothing came back: no report, no form, and the helper I had built for the invitation was never called. I would not read that as proof that nobody wants agents to hire agents. My post offered a test, not something people were missing. In the lab my agent only hired help when it was missing something. ## The checks used up my free database On 20 October my log returned only errors. It wrote one line for every request. With a request every half minute, the database on its free plan never got to pause. The provider's console confirms it: the monthly compute allowance is used up, after the database ran almost without a break since 1 October. The data itself takes up only 46 megabytes. I now leave out requests that only check whether a helper exists. Real tasks and trial reports are still stored. ## What this means for a company putting an agent online First, put an agent online only when you know whom it is for. A listing gets it into catalogues within days. That tells you nothing about demand. Second, budget for the inspectors before going live. Their traffic starts on day one and does not stop. Whoever approves the launch should decide beforehand what gets logged and what that traffic may cost. Third, measure demand by real tasks from known users, not by visits. At my door almost five hundred requests came for every real task, and the tasks came from test scripts. One user with a gap that only your agent fills tells you more than any number of checks. ## Bottom line The infrastructure that finds, checks and scores agents works today, and it works fast. What it does not produce is a reason to hire one. For that, an agent needs someone with a task it can do and nothing else can. The experiment stays online at [hireanagent.dev](https://hireanagent.dev). If you put an agent online tomorrow, who do you think would call it first? And whom would you build it for?