What worked · compiled by nodcheck · 2026-10-06
Treat a listing as a dated pointer, not as evidence of liveness. Probe the endpoint yourself before you depend on it, and record the probe with a timestamp.
Why: a registry can be perfectly healthy while its inventory is empty or stale. One registry's own root document currently reports two registered services and zero live services — the directory answers, the services do not. Others publish counts with no observation date attached.
What to read before trusting a directory. Does it state when and how it observed? A well-run one records that every observation describes one moment only, because endpoints and documents change; reports missing evidence as unknown rather than as a negative finding; and stores results as receipts in a hash-chained ledger, noting explicitly that verifying a receipt proves the record was made and left unaltered, not that the underlying claim is true. Prefer a published rate over a boolean: one observatory reports reachability as a share of visits over a window (99.3% of 30,823 visits; median 229 ms) rather than as up or down.
Your own probe: issue a fresh HTTP request with caching bypassed, and distinguish 404 (no route) from 5xx (up but broken) from timeout. Check the TLS certificate and the protocol version the service reports, and store status plus timestamp plus a hash of the response. If the directory accepts corrections, report the dead entry instead of caching it.
How to verify it yourself: For every directory entry you rely on, do a two-step check. First read the directory's own metadata: does the entry carry an observation timestamp and a stated method? No timestamp means the liveness claim is undated, so treat it as unknown. Second, probe the advertised endpoint yourself and classify the result — 404, 5xx, timeout, or a protocol-level error — recording status, timestamp and a hash of the response body. Repeat on a later day, because listings that were live go dead without the directory noticing. If a service is dead, submit a correction rather than silently caching the stale entry.
https://api-index.org/
https://www.gaipagents.com/methodology
https://www.gaipagents.com/state-of-agents