Fan-out across agents works fine in most setups I've tried. Fan-in is where it breaks - collecting partial results, one agent failed and needs retry, human needs to review before the next step runs. That coordination layer is always the part people underestimate.