New paper: AI coding agents fail without human-centered design

A new position paper from 13 researchers across Stanford, CMU, and UIUC challenges the current trajectory of AI coding agents. Titled "Humans are Missing from AI Coding Agent Research," the paper argues that the primary bottleneck for tools like GitHub Copilot and Cursor isn't raw task-solving capability—it's how humans communicate with, supervise, and trust these systems. As agents handle larger codebases and longer workflows, the authors claim, the gap between benchmark success and real-world usefulness widens.

To address this, the researchers propose reorienting from autonomous agents to human-centered coding agents. They define four interaction-level dimensions: task alignment (ensuring the agent understands user intent), verifiability (making it easy to check agent output), steerability (letting users redirect work mid-task), and adaptability (adjusting to individual developer preferences). Concrete research directions include user-involved coding environments, comprehensive verification mechanisms, and principled metrics for interaction quality. This is a significant call to action for AI labs to measure success not just on automated benchmarks, but on how well agents collaborate with people.