I think the paper allowed for a repl with variables to persist and a recursion depth for the subagents. I don't think this was implemented (though it looks like there were a few commits since I forked the repo). I've been testing out here with a larger recursion depth of 3 and a repl (and forcing the top level agent to use the subagent opaquely). I've seen an improvement in tokens consumed at the top level (though of course the haiku tokens are more used).
I think the paper allowed for a repl with variables to persist and a recursion depth for the subagents. I don't think this was implemented (though it looks like there were a few commits since I forked the repo). I've been testing out here with a larger recursion depth of 3 and a repl (and forcing the top level agent to use the subagent opaquely). I've seen an improvement in tokens consumed at the top level (though of course the haiku tokens are more used).