What happened after the tool said “success”?
A music service can accept a play request while the stream fails. That distinction matters when an agent tells someone that music is playing: the response may only establish that the service received the command.
I work on integrations around Hermes, an open-source agent framework, for the services and devices I use. An action response and a subsequent status reading answer different questions. I choose the check according to what the agent is about to claim.
Match the observation to the claim
For a changed setting, I want a fresh reading of that setting. After restarting a service, I check that the service responds through the interface I need. For this website, I compare the content served at the public URL with the intended release.
Those checks also have limits. A playback status does not prove that someone heard sound. A thermostat’s new target temperature does not establish that the room has reached it. An origin server returning a new page does not tell me which version a cache serves to visitors.
Say what is still unknown
If the confirming read fails, the agent should say that it sent the request but could not confirm the result. If it used an older reading or a fallback, the reply needs to make that clear as well.
The project does not automatically verify every possible action. The useful distinction is specific: what did the tool accept, what did the next observation establish, and what is still unknown?
Further reading: Google SRE: Monitoring Distributed Systems.