Google Researchers Propose Three-Level Proactivity Taxonomy for Coding Agents: Reactive → Scheduled → Situation-Aware, with IDQ/CGS/LL Evaluation Metrics

arXiv·medium signal

Google's Nghi Bui and Georgios Evangelopoulos argue (arXiv 2605.06717, May 7) that the field conflates proactivity with autonomy. Their three-level taxonomy — Reactive (waits for commands), Scheduled (runs on triggers), Situation-Aware (notices relevant changes, connects cross-tool signals, decides when to interrupt, carries preferences across sessions) — gives builders a concrete framework for evaluating which level their agent actually operates at. The paper introduces three evaluation targets: Insight Decision Quality, Context Grounding Score, and Learning Lift.