Constructing a string by printing things during the parsing phase is a really bad idea. Things don't necessarily happen in an intuitive order, and some productions may be applied but then rolled back based on the rest of the parse (especially in recursive descent parsing). Use the productions for what they were intended -- returning a semantic value. In this case, the productions should return the appropriate parts of the XML representation.

my $p = Parse::RecDescent->new( <<'END_GRAMMAR' );

      parse: item(s)         { "<ROOT>" . join("", @{$item[1]}) . "</R
+OOT>" }
       item: /\w+/ /\([^\)]+\)=/ data
                             { "<$item[1]>$item[3]</$item[1]>" }
       data: aggregate | string | integer
  aggregate: "(" item(s) ")" { join "", @{$item[2]} }
     string: /'([^']*)'/     { $1 }
    integer: /\d+/           { $item[1] }
END_GRAMMAR

undef $/;
print $p->parse(<DATA>);
[download]

This ignores and does not enforce the "data types" (agg, text N, int) in your markup language. But that would be easy to fix by changing the productions to:

item: /\w+/ "(agg)="           aggregate
    | /\w+/ "(text" /\d+/ ")=" string
    | /\w+/ "(int)="           int
[download]

(and of course updating the semantic actions accordingly)

Another thing to note is that in your input, there are two layers of "TITLE" tags, but your desired output has only one. You'd have to add a special-case productions to give your desired output (or post-process the XML output):