Arena Analyzes 20,840 Coding-Agent Traces to Compare Model Praise and Complaints

Arena has analyzed 20,840 traces from Agent Arena: Code across 29 models, arguing that direct user praise and complaints can become a practical signal for tracking coding-agent performance. The study, covering activity from August 1 to September 13, found that newer frontier models generally receive a more positive balance of feedback, while broken code remains the most common source of complaints.