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

This paper provides the first controlled evaluation of progressive disclosure—a technique where documents are packaged into hierarchical 'skills' that an agent loads on demand. The authors compared three navigation strategies: raw-document navigation, flat disclosure (a single index), and hierarchical disclosure (recursive routing). Using the new LOONGDOC environment, they tested these approaches across three agent harnesses and three model families on the InfiniteBench dataset, measuring performance on both single-book and multi-book (library-scale) tasks.

For single-book tasks, the benefit of progressive disclosure is highly dependent on the agent's native capabilities. Strong agents that already possess effective 'grep-like' navigation strategies gain little from pre-packaged skill sets. However, for agents that struggle with raw navigation, flat disclosure provides a meaningful performance lift.

When scaling to library-sized corpora (multiple books), the value of progressive disclosure becomes clear. Raw-document navigation collapses as the volume of text increases, while one-level flat disclosure maintains significantly higher accuracy. Crucially, the authors found that 'more is not better' regarding depth: adding a second level of hierarchical routing never improved accuracy and often caused catastrophic failure, likely because the always-loaded descriptions of the deeper hierarchy consumed too much of the agent's limited context window.

Practitioners have widely adopted progressive disclosure based on anecdotal evidence, often assuming that deeper, more granular hierarchies are superior. This study provides empirical guardrails for agent architecture, suggesting that developers should prioritize a single, well-structured flat index rather than complex, multi-level hierarchies. It clarifies that progressive disclosure is a tool for managing context limits rather than a substitute for agent intelligence.