Rho: A Purpose-Driven Workspace for Human-AI Collaborative Analysis

Abstract

Agent-assisted bioinformatics has moved the bottleneck from computing a result to judging it: whether a comparison answers the intended question, on which unit of replication, at which version of the input, and what would change the conclusion. These judgments need materials that are usually scattered across conversations, notebooks, figures, run logs and notes, while the agent's context is a transcript rather than the structure the researcher is actually reading. We present Rho, a human--AI collaborative analysis environment that organizes materials by the current scientific purpose: it selects or obtains the materials a judgment needs, declares the relations they serve, composes operable views, and makes this organization a shared state that is at once the researcher's reading interface and the agent's continuation context. We present the design in three parts: a method object---the operations that form a round-trip, the kinds of state that must stay distinct, and the protocol invariants that make each distinction checkable; a three-layer architecture that separates capability providers from the application that composes them; and five illustrative analyses on public data. In a bulk RNA-seq case the recomputed counts differed from the published counts and the interface kept the computation, the choices and the discrepancy available for inspection; a single-nucleus case makes sample-level replication visible behind a cell-level comparison. Under one configuration these analyses show how the design behaves in practice; we present the work as a design contribution rather than an evaluation, and usability, comparative benefit and the scientific validity of individual analyses remain to be evaluated.