Why did a downstream agent get worse when its own code never changed?

Because the cause isn’t always where the failure shows up. In a multi-agent system, one agent’s output is another’s input, so a routing agent that starts summarizing differently, or a retrieval step that starts returning different documents, changes what a downstream agent has to work with, even though nobody touched that downstream agent’s code. Its own commit history stays clean and its own prompt hasn’t moved, so a search scoped to just that agent finds nothing, because the search is looking in the wrong place. Attribution has to walk the dependency graph, not just one agent’s change history: check what changed for every upstream system the failing agent depends on, in the same window, and test whether that change explains the shift. The fix usually belongs upstream too, in whichever agent’s output actually moved, even though the regression was first noticed somewhere else.