Why did an error one agent made become the next step's ground truth?

Because the wrong output gets treated as fact by whatever reads it next. A wrong intermediate result is consumed downstream as input, and the step receiving it reasons correctly from a false premise, so its own output looks internally consistent even though the chain is now wrong. Every span involved is telling the truth about itself; the defect lives in the relationship between spans, not inside any one of them.

The effect compounds with depth. A controlled study of agent pipelines found output consistency falling from 100 percent at direct execution to 23.5 percent at ten hops, with a standard deviation of 22.1 at that depth, nearly as large as the mean, so read it as a direction rather than a precise number. It’s worse still when the same model plans and verifies its own work, because the verifier inherits whatever the planner already got wrong.