1,248 GitHub agent specs show safety steps missing from most

Researchers studied 1,248 workflow files across 276 GitHub projects to understand how teams write and update automated agent jobs. Most workflows share a small set of recurring shapes: a trigger, a sequence of steps, and a set of access rights. The structures are readable at a glance. The safety decisions inside them are thin. Teams add access rights without removing old ones, so privilege grows over time rather than shrinking. Most files grant agents broad access to act on code and issues. Fewer than a third include any step where a human reviews or approves what the agent did before it takes effect. Changes were small and additive. Teams patched what broke and extended what worked. For designers, the gap is not in the writing of these specs. Teams write them with care. The gap is in the human decisions the specs leave out. There is no standard place to say what the agent may spend, what it may touch, or when it must stop and wait. Budget limits, reach limits, and approval steps appear only when someone thought to add them.