Right Answer, Wrong Year: Diagnosing Stale Grounding in Search Agents

Search agents are increasingly asked questions whose correct answer depends on when the question is about, and the retrieved evidence supporting an answer is itself valid only over some stretch of time. We ask whether search agents keep track of this. Working with temporally anchored queries, we find that agents routinely commit to answers whose supporting evidence held at some earlier moment but does not hold at the moment the question asks about. We call this failure stale grounding, and observe that final-answer accuracy hides it: an agent can be right for reasons that no longer apply, and wrong for reasons it never checked. To expose it we introduce , which reconstructs, for every candidate answer and every claim supporting it, the interval over which retrieved evidence licenses that claim, and compares it against the query's temporal anchor. separates evidence that pins a claim to the anchor from evidence that merely mentions it, and reveals four recurring mechanisms: agents silently re-anchor to the present, accept undated claims as though they were current, ground on a fact that later retrieval supersedes, and stitch together hops that were never simultaneously true. We then ask whether making temporal state explicit during execution changes agent behaviour, and instantiate , an inference-time tracker that maintains validity intervals alongside the agent's own reasoning and surfaces their running intersection. Across trained and prompt-based agents, reduces stale grounding and raises accuracy while shortening rather than lengthening trajectories.