For example, in private communication Pieter suggested putting the aggregate signature data into the top of the first segwit v1+ input witness (and pop it off before evaluation of the input script) whether or not that input is participating in the aggregation or not.  This makes this canonical choice of position independent of the runtime behaviour of other scripts and also prevents the script from accessing the aggregate signature data itself, while still fitting it into the existing witness data structure. (It doesn't let us toy with the weights of aggregated signature, but I hope people will still be motivated to use taproot solely over P2WPKH based on having the option to perform aggregation.)

Being able to allow aggregation to be compatible with future script or opcode upgrades is still very difficult to design.