Is Progressive Disclosure All You Need for Long-Context Agents?

Long-document question answering usually forces a choice between loading the whole document into the context window and bolting on a separate retriever. Agentic AI suggests a broader option, giving the agent the document path and letting it decide how and what to read. Agent Skills, a standard for packaging expertise into folders an agent loads on demand, supply a ready mechanism: progressive disclosure, which exposes only what a query needs, from a short description down to the specific passages. Practitioners rapidly adopted this pattern for book-length understanding tasks, but the evidence for it has been anecdotal. We run the first controlled study of the pattern, comparing raw-document navigation and several designs of Agent Skills packs against a classical hybrid retriever across three agent harnesses and three model families on ∞Bench. On a single book, the gain depends on the harness, running large when the agent navigates the raw document poorly but near zero when a strong agent harness already locates and reads the right passages on its own. When scaling up to tasks that span many books, raw-document navigation collapses while one-level progressive disclosure degrades more slowly and pulls ahead.

Introduction. Long-document question answering (QA) forces a choice between two well-studied options: load the entire document into the context window and lean on the model’s long-context attention, or retrieve a small set of chunks with an external retriever. Neither option wins outright (Li et al., 2024, 2025), and on long inputs the effective context size falls well short of the advertised window (Liu et al., 2024; Hsieh et al., 2024). Agentic AI reframes the choice: give the agent the document and let it decide what to read and when. Claude Code made exactly this move, dropping the retrieval index its early versions used in favor of letting the agent search the code base on demand (Cherny and Wu, 2025). But its engineers report the switch from experience, not a controlled comparison. Agent Skills turn that option into a mechanism.

Discussion / Conclusion. We ran the first controlled study of progressive disclosure for agentic long-document reasoning, comparing raw-document navigation, flat and hierarchical skill packs, and a hybrid retriever on ∞Bench across three harnesses and three model families. Progressive disclosure buys context, not intelligence: on a single book a strong agent already greps the passages it needs, so disclosure adds nothing, but once the corpus outgrows what native navigation can hold, flat disclosure becomes decisive. A second routing level never reproduces this gain and sometimes breaks accuracy. For packaging book-length material the guidance is concrete: package a book as one skill layered progressive disclosure, not as a multiple parallel packages of child skills with always-loaded description.