Coding agents need proactivity, not just autonomy, says new paper

A new position paper from researchers Nghi D. Q. Bui and Georgios Evangelopoulos argues that the next generation of coding agents must evolve from autonomous reactive tools to proactive, long-horizon assistants. The paper, submitted on May 7, 2026, critiques current agents that simply edit repositories, open pull requests, or respond to webhooks. Instead, the authors propose that agents should notice relevant changes before the developer asks, connect signals across multiple tools, decide when to interrupt, and carry preferences across sessions. They draw on mixed-initiative interaction principles to distinguish proactivity from autonomy.

To ground this shift, the paper introduces a three-level taxonomy of proactivity: Reactive (responds only to explicit triggers), Scheduled (runs at predefined times), and Situation Aware (acts based on context and anticipatory reasoning). The authors also outline five practical acceptance criteria for proactive agents and three new evaluation metrics: Insight Decision Quality (IDQ) – how well the agent decides what matters next; Context Grounding Score (CGS) – how accurately it connects evidence from the codebase and tools; and Learning Lift – how it adapts after feedback. The paper is intended to guide future research and tooling, emphasizing that unsolicited agent behavior must be judged by its usefulness, not merely its activity.