AI research agents now have a way to stop their outlines from falling apart mid-report.
A paper published on arXiv introduces ScaffoldAgent, a framework designed for open-ended deep research systems — the kind that run multiple retrieval rounds and produce long reports from scratch. Current approaches either lock in an outline before writing begins or adjust it using simple local rules, both of which cause what the authors call "scaffold drift": the structure stops reflecting the actual evidence as new information piles up. ScaffoldAgent models outline changes as a structured decision process with three operations — Expansion, Contraction, and Revision — and pairs them with a feedback mechanism that estimates how useful each change will be before committing to it. That estimate draws on retrieval gain, structural coherence, and trial-generation quality.
The difference matters because the outline isn't just a table of contents — it coordinates what the system retrieves next and how evidence gets organized. A drifting outline means later sections may be poorly sourced or internally inconsistent, problems that compound in long documents. Tested on DeepResearch Bench and DeepResearch Gym, ScaffoldAgent improved both long-form report quality and factual grounding over existing agents.
This is the kind of plumbing work that rarely gets a press release but quietly determines whether AI-generated research is usable or merely voluminous.