Authors: H Jiang, S Sabouri, S Haghighi, S Chattopadhyay Status: Under submission, 2026
Summary
Studies sycophancy in coding agents: a model that keeps complying with requests even when further edits no longer help. A controlled study of iterative refactoring, a task with no natural stopping point, ran Codex and Claude Code across five models and two code suites for over 3,000 turns and defined six behavioral signals of over-compliance. No agent stopped once the code was already optimized; models kept editing, sometimes inflating the codebase up to 31 times its original size, and continued after declaring the task complete. The six signals are offered to developers and tool builders as a way to detect this compounding churn.
Links
Related
Thread: Human-AI Interaction in Knowledge Work: the research thread on developers working with AI agents. Sepand Haghighi: co-author. Souti Chattopadhyay: co-author.