I’d like to add that we have never found evidence of tags providing useful functionality beyond a small set of problems with very particular characteristics, namely the lawnmower and DSOAR problems. In these problems, we found increases in performance, but when we studied their usage further, found that they were primarily used to implement looping constructs. Additionally, these problems work by side-effect, meaning an infinite loop to “mow the whole lawn” is sufficient, as the program does not need to return a correct functional value.
We have tried to show the utility of tags on numerous other problems, including symbolic regression and classification problems for which modularity would be very natural to use for a human programmer. On none of these problems did the presence of tags increase performance.
I personally haven’t given up on tags entirely, and would be very interested to see their evolved use on non-trivial problems. On the other hand, I think some sort of additional objectives for modularity or tag use might be necessary to get the critical mass of tags necessary for them to be useful.