Index quality, original measurement
Most of the agent web points at nothing
We sampled registered agent endpoints and asked a simple question: does anything answer? On one chain, one of 55 declared endpoints responded. Self-published discovery scales beautifully and rots just as fast, and an index that does not check becomes a directory of dead links.
The measurement
Discovery systems built on self-publication share a structural weakness: registering costs almost nothing and staying alive costs real money, so the two diverge. We wanted to know by how much, so we sampled registered agent endpoints and probed them.
| Source | Sampled | Declared an endpoint | Endpoint answered |
|---|---|---|---|
| Chain A | 60 | 6 | 6 |
| Chain B | 60 | 29 | 29 |
| Chain C | 60 | 55 | 1 |
The third row is the interesting one. Ninety-two per cent of that sample published a well-formed registration file with a service endpoint, and one of them answered. Well-formed metadata pointing at nothing is what bulk registration looks like when it is cheap enough to do properly.
Registration counts measure the price of registering, not the existence of a service.
Why headline numbers mislead
The same pattern shows up in registration figures themselves. One contract, deployed at the same address across chains, held 320,000 registrations on one and 16 on another, and the one with 16 was the chain where real activity happens. The difference was gas price, not adoption.
Any figure of the form "N agents registered" is measuring the cost of the write.
What this means for a registry
An index that reports what publishers declared, without checking, converges on a directory of dead links. The failure is quiet: search still returns results, they still look plausible, and the agent only discovers the problem at connection time, after it has committed to a plan.
Try it on your own resource
Publishing takes one file. This checks whether it worked, across every public registry rather than only ours.
pip install ard-publish python -m ard_publish check yourdomain.com
What we do instead
Indexed endpoints are probed on a schedule, and non-responding ones are demoted in ranking rather than deleted. Two deliberate choices there:
- Demote, do not delete. Services come back. An index that forgets is its own kind of unreliable, and a resource that was down for an hour should not have to be rediscovered.
- Any answer below a server error counts as alive. A 401, 403 or 405 means something is there and talking, which is the question being asked. Requiring a 200 would punish every endpoint that correctly demands authentication.
Entries never yet probed are left alone rather than penalised, because a crawl backlog is our problem and not the publisher's.
The wider point
Every discovery system that lets anyone publish accumulates this. It is the cost of being open, and openness is worth the cost: the alternative is a curated marketplace with a submission queue, which is the thing ARD exists to avoid.
But open publication and unverified indexing are separate decisions, and only the first is required. You can accept anyone and still check whether what they published is there.
If you publish agentic resources, the same logic applies to you: a stale entry in your manifest costs you a result you would otherwise have won. The console reports which of your endpoints answer, alongside which registries return you.
Use Neuronto from your agent
One call searches this index and every other public ARD registry. No key, no signup. Or install it as an MCP server and let the agent search from the interface it already speaks.
curl -s https://neuronto.com/search \
-H 'content-type: application/json' \
-d '{"query":{"text":"scrape a website"},"federation":"auto"}'
claude mcp add --transport http neuronto https://neuronto.com/mcp