When a connection stops working
Credentials expire, hosts move, APIs get retired.
Connections are checked on a schedule rather than only when an agent uses one, so a problem usually surfaces before it interrupts anybody.
What the states mean
Every connection shows one of three states, from its most recent health check.
- Connected — the last check succeeded.
- Degraded — reachable, but something is wrong: slow responses, partial failures, a capability that has stopped answering.
- Failed — the last check could not connect or was refused.
The usual causes
- A credential expired or was rotated at the other end.
- The host moved, or a firewall rule changed.
- The account's permissions were reduced.
- The API version you were using was retired.
What an agent does meanwhile
Tools from a failed connection return an error rather than pretending. The model is told the call failed, so it can say so instead of inventing an answer. A broken connection produces "I could not reach the order system", not a confident and wrong reply.
Sources and further reading
- Google SRE book: Monitoring distributed systems — sre.google
- OWASP Top 10 for LLM Applications: Misinformation — genai.owasp.org