Haystack's agent is indeed using the approach suggested in the ReAct paper