Mesa has support for reactive programming via mesa.experimental.mesa_signals. Upon reviewing this code again, there may be room for some improvement.

1. Replace AttributeDict with a (frozen) dataclass.
2. The current implementation is designed around the use of descriptors. What about rewriting this to use the property factory pattern recently introduced in property layers?

While exploring these changes, it is crucial to explicitly assess their performance implications. When signals were added, it was seen as a potential building block for much more flexible data collections. Early benchmarks yielded mixed results from a runtime perspective. Some stuff became faster, other stuff became slower because of the signal overhead.