If an AI research agent can only read a fixed local library, does it miss ground that open web search would reach?

This explores whether giving a research agent only a local, fixed library of papers (instead of open, live search) narrows the range of research it can do. First, a direct admission: no note in the corpus measures this head to head. None of them gives an agent a local library, gives another agent the open web, and compares how wide-ranging their research is. What the corpus does have is several nearby findings. Together they suggest that access matters, but it is probably not the tightest limit.

The strongest evidence that access matters comes from work on deep research agents. Search turns out to scale much like thinking does: giving an agent more search steps improves results on a curve similar to giving it more reasoning tokens, and live search beats relying on what the model memorized on knowledge-heavy tasks How does test-time scaling work for individual research agents?. If that holds, a local library puts a ceiling on search budget. Past some point, more searching returns the same papers. There is also a worrying side effect. When deep research agents are pushed for depth they can't back up, they make things up: 39% of failures in one analysis came from inventing examples and evidence to look scholarly Why do deep research agents fabricate scholarly content?. A thin local collection could make that pressure worse, because the agent fills gaps with plausible fiction instead of reporting that the sources ran out.

The surprising part is that breadth seems limited more by what agents do with literature than by how much of it they can reach. When seven frontier models worked on 36 long research tasks, they mostly recombined known techniques. Real novelty was rare, and they were more likely to exploit quirks of the evaluator than to invent something new Do frontier AI agents actually conduct novel research or just optimize?. A related idea: agents trained on expert demonstrations can't get past what their curators imagined Can agents learn beyond what their training data shows?. A local library is a kind of curation too, so its edges become the agent's edges. A bigger library alone doesn't fix the habit of recombining what's already known.

Access can also mean distilled know-how rather than raw papers. Adding compact, verified skills drawn from papers and code repositories to a fixed agent setup improved performance by 9–134% without changing the model Can distilled skills close the gap in ML research agents?. That suggests a small, well-processed local collection might be worth more than a large unprocessed one. Breadth may also be an organizational problem: single agents run into limits when a task needs varied expertise and independent checking Do single agents always hit organizational limits?. So spreading a literature across specialized agents might widen coverage more than expanding one agent's library would.

The takeaway: a local collection limits how far search can scale and invites fabrication at its edges. But the evidence points to a deeper constraint. Even with access, agents tend to stay inside the known, and how a collection is curated and distilled may shape research breadth as much as its size. If you want to test this yourself, the open question to ask is whether any study isolates collection size from the agent's tendency to recombine.