It sounds like we've had similar experiences. I converted a handful of ad hoc "state machines" in a telephony system to FA. The original code was loaded with ex

It sounds like we've had similar experiences. I converted a handful of ad hoc "state machines" in a telephony system to FA. The original code was loaded with exceptions and special cases (and bugs). Converting to FA required detailed analysis of the existing system to extract the distinct events and states. I used a tool I wrote [1] to generate C state tables and an event loop from a description of the states+events. While the state machines that came out of the analysis were a big improvement, they still weren't perfect. We came across a few bugs later on. Using a strict representation of the state machines was a huge advantage in debugging and maintenance. When everything is moving through the same event loop, it's easy to add logging in the form "[timestamp] state A: event X -> state B". Adding a new state or event to the machine (to fix a bug or add a new feature) is much easier when everything is in one place. In a couple of cases, the state machine started to become large and unwieldy. Usually the cause was that there would be a couple of "meta" states: we'd have states A, B, and C, but then we'd need to represent "states B and C when X is true; states A and B when Y is true". Pulling X and Y into a separate machine simplified the parent state machine -- every X and Y moved into the child divides the number of states in the parent by as much as half. Then the parent can inject events into the child, and the child can fire events back to the parent. [1] http://bstpierre.org/Projects/smc/ http://bstpierre.org/Projects/smc/