I implemented something similar at work for a phone handling system. We had a "workflow" system that let clients script how calls should be handled. All side ef

I implemented something similar at work for a phone handling system. We had a "workflow" system that let clients script how calls should be handled. All side effects that the script was allowed to make happened through an API with functionality like reading DTMF input, playing media, adding another call leg etc. Like the article says, we would record a log of all the interactions with the outside world. The lower level parts of the system had the capability of transferring an in-progress call from one server to another. SIP and RTP traffic would be moved to another server, and finally the workflow log was sent. The server that took over the call would then replay the script and arrive at the same execution state that the first server had before the call was moved. Workflow authors didn't have to care about any of this. To them it looked like the call started and ended on the same server.