AI Agents for CRISPR Diagnostics: From Automation to Bounded Autonomy

Abstract

Recent years have witnessed rapid advances in clustered regularly interspaced short palindromic repeats (CRISPR) diagnostics, artificial intelligence (AI), and self-driving laboratories for automation. CRISPR diagnostics have emerged as a promising strategy for earlier disease detection, with Cas12a and Cas13a targeting DNA and RNA, respectively, and converting that recognition into collateral reporter cleavage. In parallel, different subsets of AI have advanced rapidly, with machine learning (ML) interpreting biosensor readouts, and large language models (LLMs) and AI agents applied to automation. Laboratory automation and self-driving platforms reduce the workload of human experimentalists and enable scale-up, having run long unattended campaigns in materials discovery. However, how these advances in AI, CRISPR diagnostics and automation could be integrated to enable a future of autonomous biosensing remains unresolved. In this perspective, we evaluate CRISPR diagnostics as autonomous systems, where analytical metrics alone cannot ensure robustness, failure-mode predictability, recoverability and auditability. We then propose a bounded-autonomy layer in which multimodal evidence, scored by ML, gates a Sense, quality control (QC), Decide, Act and Report loop, with guardrailed LLM agents invoked only when QC flags a run, selecting pre-validated actions under deterministic validation. By tying detectable failure modes to validated recovery actions, we outline a path toward bounded autonomy in CRISPR diagnostics, where a device unable to act safely abstains and escalates.