- The paper shows that progressive disclosure provides little benefit when agents can navigate a single document effectively, but substantially improves long-document QA as the corpus grows to twenty books.
- The paper finds that flat, one-level skill indexes consistently outperform hierarchical disclosure, which can sharply reduce accuracy by loading too many chunk descriptions into context.
- The paper demonstrates a practical cost advantage at library scale: flat disclosure nearly doubles English open-QA accuracy over raw navigation while using about half the tokens and cost at twenty books.
This paper presents the first controlled study of progressive disclosure—the Agent Skills pattern of exposing only what a query needs—as a mechanism for long-document question answering, comparing it against raw-document navigation and a classical hybrid retriever (2607.17598). The authors build two skill-pack designs over a single fixed chunk set derived from the widely adopted book-to-skill recipe, run them through three agent harnesses (Codex, Pi, Claude-Code) and three model families (gpt-5.4-mini, qwen3.6-27b, claude-haiku-4.5) on InfiniteBench book-QA subsets, and scale from one book to bundles of twenty. Their central finding is that progressive disclosure "buys context, not intelligence": it is redundant when a strong agent already locates passages on its own, but becomes decisive once the corpus outgrows native navigation—and one routing level is enough.
Motivation and framing
Long-document QA has historically forced a choice between loading the full document into the context window and bolting on an external retriever, with neither option winning outright: effective context size lags well behind advertised windows due to position-dependent degradation, and the long-context-versus-RAG literature reaches contradictory conclusions. Agentic AI reframes this as giving the agent the document path and letting it decide what to read. Claude Code's removal of its retrieval index in favor of on-demand search is cited as the motivating precedent, but that switch was justified by engineering experience rather than controlled measurement.
Agent Skills supply a concrete mechanism for agentic reading: a directory rooted at a SKILL.md file whose always-in-context description gates discovery, whose body activates on match, and whose bundled files execute on demand. Practitioner literature promotes this as a near-replacement for retrieval, including packaging whole books so agents read only relevant chapters—but with token-savings claims reported under no accuracy control. The authors identify two confounded design axes in the specification that prior work never separates: how deep disclosure recurses, and where per-chunk index metadata physically lives (always-loaded description context versus load-on-activation body context).
Methodology
The study isolates three routes to the same passage:
- Raw navigation: the agent receives the bare document in a sandboxed file system with no prepared structure.
- Flat disclosure: one SKILL.md whose always-loaded description describes the whole book; its body indexes chunk files under
references/ with per-chunk descriptions that cost context only after activation.
- Hierarchical disclosure: each chunk is itself a skill with an always-loaded description; a meta-router routes among children, making reading one step shorter but paying an always-on context tax for every chunk description.
Both packs share one chunk set—split along chapter headings or ~4000-word paragraph boundaries—with LLM-written descriptions pairing summaries with named-entity lists, generated at temperature 0 with fixed seeds. A hybrid-RAG baseline (BM25 + BGE-M3 dense retrieval, reciprocal rank fusion, BGE cross-encoder reranking) serves as the classical reference point.
LoongDoc environment
Because static benchmarks cannot observe navigation or host discoverable skill packs, the authors rebuild InfiniteBench into LoongDoc, an interactive environment built on BenchFlow over the Agent Client Protocol. Each question becomes a task where the agent reads, greps, and opens files at will; every rollout logs tool calls, opened files, and per-step token usage, and the deterministic verifier emits a scalar reward per trajectory. The verifier matches the official InfiniteBench scoring exactly—regex option extraction for multiple choice, normalized match and token-level F1 for open QA—with no answer-extraction model in between. At library scale, K books are staged with a corpus-index.md and label-prefixed question identifiers, changing only how many files are written, not the interface.
Single-book results
Three findings emerge from the single-book grid:
Disclosure helps where native navigation is weak. Under Pi and Claude-Code, flat matches or exceeds raw in nearly every cell—for example, Pi gpt-5.4-mini En.MC rises from 0.8851 to 0.9126. Under Codex, however, all three approaches tie within error across all subsets. Trajectory inspection explains why: the bare Codex agent does not read linearly but greps the raw text for entities named in each question and reads only matched passages, reconstructing locate-then-read retrieval on the fly. For such agents, a pre-cut skill buys controllability of the retrieval path, not accuracy.
Depth never pays and can hurt. No cell rewards hierarchical over flat. On Pi it collapses outright: En.MC drops from 0.9126 to 0.6398 on gpt-5.4-mini, and Zh.QA falls from 0.7479 to 0.3890 on qwen3.6-27b—drops far outside cell variance, attributed to always-loaded child descriptions saturating the router's context before it commits to a chunk.
The gain is not retrieval in disguise. Hybrid-RAG trails both raw and flat on qwen3.6-27b across all three subsets, with the widest gaps on open QA, indicating that routing through an in-skill index outscores the classical rerank-and-feed pipeline.
Library-scale results
Scaling from K=1 to K=20 books inverts the picture. Under raw navigation, even Codex degrades sharply: En.QA more than halves and Zh.QA collapses to near zero ($0.043$ at K=20), because the same agent that matched the skill pack on one book cannot reliably find the target among twenty. Flat disclosure degrades more slowly and pulls ahead—at K=20 on En.QA it holds $0.462$ against raw's $0.257$, clearing one standard error, and it leads all six En.QA cells across both harnesses and both models, which the authors identify as the most robust effect in their tables. Cost accounting sharpens this: at K=20, raw burns 68.3M tokens per question (~$52 uncached) for the worst accuracy, while flat reaches nearly double the accuracy at roughly half the tokens and cost. On the cost–accuracy Pareto frontier, every configuration recedes down and to the right asKK{=}10K=5K{=}11K≥10.
Hierarchical fails to reproduce the rescue on open QA: always-loading every chunk's description recreates the very context pressure disclosure is meant to relieve once twenty books inflate the always-on budget. An appendix on the weaker Pi harness adds nuance: depth is a scale- and task-specific effect rather than a uniform cost, since hierarchical recovers on Pi's large-corpus open-QA cells while remaining worst on En.MC there.
Limitations
The most serious threat is a pre-training confound on En.MC: the multiple-choice books are canonical English novels models have likely memorized. Two traces support this—even at K=12 the raw agent answers correctly in too few tool calls to have read twenty books, and single-book trajectories show the agent recognizing renamed canonical works—which blunts any disclosure signal on that subset. The scaling axis is sampled coarsely at K=13 within one benchmark family, leaving transfer to non-narrative corpora untested. Gains are task- and language-specific: sharpest on English open QA and absent or negative on Chinese, where base-model capability rather than the skill pack sets the ceiling. The chunking and description recipe is held fixed at one procedure, and sample sizes inherited from InfiniteBench leave several comparisons within one standard error. The authors note the pipeline would accept any long-context benchmark with local gold answers; NovelQA was excluded because its labels are withheld behind a third-party scoring platform.
Conclusion
This study converts an anecdote-driven engineering practice into measured evidence. Progressive disclosure functions as a harness-dependent context-scaling tool: negligible for a single book under a strong navigating agent, decisive for a library too large to traverse by reading, and best delivered as a single flat index rather than recursive child skills. The practical guidance is concrete—one skill per book with layered disclosure, not parallel child packages with always-loaded descriptions—and the open questions are equally specific: whether the effect transfers beyond narrative corpora, where precisely between K=14 and K=15 the crossover occurs, and how alternative chunk granularities shift the balance.