The next version of XState will be much more ergonomic, with a reduced API surface area, lower learning curve, and much easier for devs (and agents) to author.
But at the same time, frontier models are very good at writing XState.
It feels bad because I know your team was working so hard pre AI to make state machines more accessible and bring such a powerful concept to the world.
I know it didn't pan out exactly, but I really want durable execution state machines that exceed workflows one day.
I keep thinking to myself can I enforce my AI layer to have some deterministic durable state machine running it and I'm not deep enough on them from a creator/author perspective to answer that question like you, but keep us posted on the changes.
But at the same time, frontier models are very good at writing XState.