# 长上下文智能体：渐进式披露并非万能

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-07-24 15:56
- AIHOT 分数：37
- AIHOT 链接：https://aihot.virxact.com/items/cmryoavcj03cprolgrm9tdn4m
- 原文链接：https://x.com/rohanpaul_ai/status/2080562628431352264

## AI 摘要

一项新研究对比了原始导航、扁平技能与深层技能层级在长上下文智能体中的效果。在单本书场景下，强智能体本身搜索能力已足够，扁平技能几乎无准确率提升；但在20本书的大型知识库中，扁平技能将英文开放问题准确率从0.257提升至0.462。研究表明，技能索引并非让模型更聪明，而是在语料库过大时帮助节省上下文，且单层路由在大规模文档库中比深层层级更有效。

## 正文

Agent skills do not make models smarter； they help models avoid wasting context while finding the right text.

i.e. agent skills are not a universal replacement for search or stronger reasoning.

They matter most when the knowledge base becomes too large for an agent to browse efficiently， and the simplest 1-level layout worked best overall.

When an agent can search 1 book well， progressive disclosure adds little， but it helps across large libraries.

A single routing layer preserves context better than deeper skill hierarchies across large document libraries.

The team compared raw navigation， flat and deeper skill layouts， and a standard retriever across several agent setups.

On 1 book， strong agents already searched well， so the flat skill brought almost no accuracy gain.

With 20 books， English open-question accuracy rose from 0.257 under raw navigation to 0.462 with the flat skill.

Strong agents gain little from prepared skill indexes until the corpus becomes too large to navigate.

---

- arxiv. org/abs/2607.17598

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