Does progressive disclosure help agents navigate large document collections?

The paper's central claim is that progressive disclosure, which exposes only what a query needs "from a short description down to the specific passages," "buys context, not intelligence." It reports the first controlled study of the pattern, comparing raw-document navigation, flat and hierarchical Agent Skills packs, and a hybrid retriever on ∞Bench across three agent harnesses and three model families. On a single book the gain depends on the harness: it "runs large" when the agent navigates the raw document poorly and is "near zero" when a strong harness already locates and reads the right passages on its own. When tasks span many books, raw-document navigation collapses while one-level progressive disclosure degrades more slowly and pulls ahead.

The reasoning the authors give is about headroom. Loading a whole document and retrieving chunks "neither wins outright," and effective context falls short of the advertised window, so letting the agent decide what to read is attractive. But on one book "a strong agent already greps the passages it needs," so a disclosure layer has nothing left to add. Only "once the corpus outgrows what native navigation can hold" does flat disclosure become "decisive." The value of the layer therefore tracks the gap between corpus size and what the agent can navigate unaided. The paper also finds that "a second routing level never reproduces this gain and sometimes breaks accuracy," and its packaging guidance follows: one skill with layered disclosure, not several parallel child-skill packages whose descriptions are always loaded.

This qualifies the on-demand-lookup family in the vault. Can LLMs read long documents like humans do? builds a fixed pipeline of gist memories with lookup of the original passages and beats retrieval baselines. This paper hands the layering to the agent and finds the benefit conditional on harness and corpus size rather than general. It also sits beside Can delegation teach models to manage context more actively?, another agent-side way to manage context, though the excerpt says nothing about subagents. Both suggest context is now managed by the agent, but this paper adds that the management only pays where the agent's own navigation runs out.

The excerpt does not establish several things. It does not name the three harnesses or model families, say how many books mark the point where raw navigation collapses, or give any effect sizes. It lists the hybrid retriever as a comparison but reports no result against it. It offers no explanation for why a second routing level hurts. The introduction's Claude Code example is, in the authors' own words, an engineers' report "from experience, not a controlled comparison." The defensible reading is conditional: check whether a strong harness already navigates the raw document before adding structure, and if the corpus is large enough to need it, keep the structure one level deep.