Flat skill packs lift 20-book QA accuracy from 0.26 to 0.46 — a second routing level erases the gain

Notes on Is Progressive Disclosure All You Need for Long-Context Agents? (arXiv:2607.17598) — Yifeng He, Yin Zhao, Jicheng Wang, Hao Chen · July 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

Agent Skills packs — the folder-based standard for handing an agent on-demand expertise — expose only a short description until a task matches it, then load an indexed body, then the specific passages: that's progressive disclosure. The first controlled test of the idea pits raw-document navigation against flat and hierarchical skill-pack designs, plus a classical hybrid retriever, across three agent harnesses (Codex, Pi, Claude-Code) and three model families. The benchmark itself is new: LoongDoc turns the ∞Bench long-document set into a live environment where agents navigate a sandboxed file system instead of a fixed context window, testing single books and libraries of 5, 10, and 20 books stitched together.

The payoff hinges entirely on how well the underlying harness already navigates raw text. On a single book, flat disclosure lifts the weaker navigators — Pi, Claude-Code — but adds nothing under Codex, whose bare agent already greps for the entities a question names and reads only the matched passages. Push one level deeper into a hierarchical pack, where every chunk becomes its own skill under a meta-router, and the picture gets worse: one Pi/model pairing sees multiple-choice accuracy collapse from 0.91 to 0.64. Once the corpus grows to 20 books, the ranking flips — raw navigation's accuracy on English open-book QA falls to 0.26 while flat disclosure holds at 0.46, a rescue that reproduces across two harnesses and two model families. The hierarchical pack never reproduces it.