AI Coding Agents Lack Human-Centered Design, Paper Argues

A new position paper on arXiv (2608.12355) argues that AI coding agent research has focused too heavily on autonomous task-solving, neglecting the human-agent interaction loop. The authors identify four core dimensions—task alignment, verifiability, steerability, and adaptability—as critical for effective collaboration between humans and AI agents. They propose a reorientation toward human-centered coding agents and outline research directions including user-involved coding environments. The paper highlights that as agents improve, the primary bottleneck shifts from capability to communication, supervision, and trust.