Melanie Warrick used an AI Engineer demo of Ziggy's ice cream delivery to pitch Temporal's durable execution for human-in-the-loop agents where the risk is remote signal injection, not local crash.
The setup runs three looping agents, fleet and customer on Google's 80K and dispatch on LangGraph, with Temporal workflows and activities persisting state and queueing human signals.
Warrick showed a wait condition that parks a single order workflow while the rest keeps running, then a signal that resumes it and survives a killed worker by replaying event history.
The security promise is that pause, yet the demo never showed who can send that signal, how it is authenticated, or what timeout stops an order waiting days for a human.
In production Temporal already underpins large systems like Coinbase trading, payments and settlement flows, but that scale does not prove fine-grained authz for agent interrupts.