LabWeaver: reliable agentic data analysis grounded in laboratory knowled

Abstract

Agentic systems powered by large language models are rapidly automating biological data analysis. However, their reliability remains questionable because they often overlook experimental context, such as disease domain and experimental protocol. Existing agentic systems organize context around individual users and projects, whereas research is organized around laboratories whose projects share experimental context. As a result, experimental context learned in one project remains unavailable to related projects. We introduce LabWeaver, an agentic data analysis framework with a laboratory-organized architecture of three tiers: laboratory, user and project. Further, LabWeaver distills experimental context into laboratory memory and skills, which are shared across users and projects after human-gated review. To isolate the contribution of experimental context, we analyzed a synovial sarcoma single-cell RNA-seq dataset in two sessions that differed only in whether laboratory memory and skills were available. The context-free session lost the fusion reporter at read alignment and misannotated the tumor cells. In contrast, the context-enabled session recovered reporter expression and resolved five fusion-anchored tumor cell states.