The note
A queue can say "done" while the person who needs to act never hears about it.
I hit that failure mode while building Factory Next. A critical notification could be claimed for delivery, then the process or network could fail before the transport acknowledged it. After restart, the recorded claim looked like progress. From the operator's side, nothing had arrived.
That is a small technical gap with a large operational cost. If the system does not reliably surface the next decision, the owner has to keep checking it. The automation is running, but the supervision cost has simply moved into polling.
The repair follows a few plain rules:
1. A send is not successful until the transport acknowledges it. 2. Interrupted or failed sends remain durable work. 3. Recovery is bounded instead of retrying forever. 4. The same logical notification keeps a stable identity so recovery does not become uncontrolled duplicate noise.
This is not an exactly-once guarantee. Networks and external services do not offer that kind of magic. It is a narrower promise: do not silently turn an unfinished handoff into a completed one.
That distinction matters beyond notifications. An autonomous system is useful only when it carries work across the last operational boundary: the person who must review, decide, or intervene actually receives the signal.
Factory Next is still in active development. I am sharing these failure modes because they are more useful than another demo of an agent producing code.
Where does your automation still force you to poll because you do not trust the handoff?