arxivcs.SEcs.AI2026-06-26
Govern the Repository, Not the Agent: Measuring Ecosystem-Level Risk in AI-Native Software
Autonomous coding agents now open and merge pull requests in shared repositories at scale, and the field evaluates them the way it has always evaluated components, one agent at a time, on isolated benchmark tasks. Yet agents that each pass their own tests still leave repositories…