Notebook-based alignment of human and agentic reasoning in single-cell biology

Abstract

Agentic AI is increasingly deployed on complex problems, often using chain-of-thought prompting to ground predictions in stepwise reasoning. In biomedical research, assistive agents could make this reasoning accessible to human scientists: for example, intermediate conclusions could be critically evaluated based on shared reasoning and sycophancy - the tendency to affirm a users claims regardless of their validity - could be mitigated. However, this interaction requires an alignment of reasoning between agents and humans that is difficult when multiple data modalities are considered, for example in single-cell biology, where human scientists reason through computational notebooks that contain code, results, and text. To address this issue, we developed kai, an agentic AI that iteratively generates analyses in computational notebooks. We find that kai can flexibly use multiple tools in sequence to solve complex cell type annotation problems. Compared with one-shot generation, kai demonstrates improved robustness against code errors and sycophancy, improved reasoning, and the ability to formulate and address questions on data. kais design is model-agnostic and, therefore, scales directly with advancements in large language models.