Atelier: A Human-Owned Protocol for Auditable AI-Assisted Research
Yuchong Geng ⋅ YUCHEN LU
Abstract
Atelier is a file-backed protocol that routes AI-assisted research through three layers, augmenting the researcher rather than replacing them. The researcher remains principal, a manager agent scopes work and resolves agent reports, and worker or critic sub-agents operate inside bounded packets. Every consequential exchange is recorded in durable markdown files, producing a reconstructible chain from intent to claim. A pilot study with eight runs over two sweeps shows Atelier produces structurally better orientation and lifecycle outputs than a freeform baseline, with the gain holding across four model and effort settings.
Chat is not available.
Successful Page Load