PersonaCodeBench is an evaluation benchmark for coding agents that focuses on their ability to satisfy persona‑driven requirements during multi‑turn interactions. The benchmark translates user characteristics—such as role, technical skill, interaction preference, and practical constraints—into verifiable obligations, covering implementation, permission, testing, explanation, and delivery. At runtime a controller simulates user responses, decides when to disclose requirements, and uses online requirement evaluation to give immediate feedback and advance the session. All requirements have their content, disclosure conditions, and acceptance criteria fixed before evaluation. The system includes a completion verifier for native correctness and a personalization verifier for persona compliance; a task passes only if both succeed and the declared budget is respected. We evaluate all major frontier coding agents, publish a comprehensive leaderboard, and conduct detailed experiments that demonstrate the benchmark’s reproducibility and effectiveness. PersonaCodeBench provides a unified platform for assessing whether coding agents can meet explicit, personalized user expectations alongside coding objectives.
Review: By formalizing user personas as verifiable obligations, this benchmark fills a gap in interactive coding evaluation and offers clear metrics for future agent development.