Search infrastructure has always leaned on a limit nobody designed: people get bored. You try a couple of searches, skim the results, give up, and go do something else. Autocomplete, saved copies of popular results, the dead overnight hours when ops teams do maintenance: all of it was tuned to how fast real humans lose patience.
AI agents (programs that search and act on their own) don't get bored. A new study watched them work and found they keep rerunning searches they've already tried. Tweaking the search engine's settings didn't reliably fix it. If ranking stays the same, the same search just hands back the same pages.
So the search system will probably have to supply the limit boredom used to. It needs to catch repeats, reuse results it already sent, and plan for traffic that never sleeps. Better to start measuring this now than find out at 3 AM.
~1 million: agent search sessions in the Agentic Search Queryset (PLOS One, Oct 21), generated by free, public AI agents answering test questions.
186: the most searches one agent ran for a single question. Two other question sets topped out at 131 and 77.
+401%: the worst case for extra searches, roughly 5 times what the starting question called for.
2–3: rewordings a typical person tries before giving up.
+70%: more searches when the agent moved from a 3-billion to a 7-billion-parameter model (parameters are a rough measure of AI size).
Small effect: right answers leaned toward shorter, less repetitive sessions, but the link was weak and didn't always show up.
Caveats: test questions, not real requests. Public agents only, no company data. The human comparison uses older search records, not a side-by-side test.