EvoCode-Bench is a multi-turn coding benchmark with 26 tasks spanning 227 sequential rounds, designed to evaluate AI coding agents the way developers actually use them — iteratively, with evolving requirements and a persistent workspace. Unlike single-turn benchmarks, it runs cumulative test suites after each round, checking all prior requirements alongside new ones. Key findings: single-turn scores overstate reliability by 1.4–4x compared to multi-turn performance; pass rates drop from 46.7% at Round 1 to 7.7% by Round 10; regressions — not missing features — are the dominant failure mode for top-tier models; and agents that maintained a persistent requirements document more than doubled their success rates. Model rankings also shift significantly under multi-turn pressure.
