One agent doing one job is straightforward. The trouble starts when you chain them: research feeds writing, writing feeds critique, and somewhere in the middle the thing you cared about drops out. Three ways it goes wrong: - Context decays at every handoff until step four is working from a summary of a summary - Early errors get laundered into confident output later - Nobody can tell you which step caused the bad result You'll leave with patterns for passing state between steps, a way to trace a bad output back to its source, and an honest test for when a chain beats one well-briefed agent.