# Concurrency Failure Patterns

This work presents the first systematic study of native dynamic concurrency in Codex, Claude Code, and Kimi Code, running controlled concurrent-versus-sequential comparisons on 354 tasks from four long-horizon benchmarks and one bounded benchmark for 2,124 executions; concurrency helps most on long-horizon tasks and little or negatively on bounded ones, raises token use to 1.41–3.31 times sequential, and yields a taxonomy of four main categories, 13 subcategories, and 28 concurrency failure patterns.

This work presents the first systematic study of native dynamic concurrency in Codex, Claude Code, and Kimi Code, running controlled concurrent-versus-sequential comparisons on 354 tasks from four long-horizon benchmarks and one bounded benchmark for 2,124 executions; concurrency helps most on long-horizon tasks and little or negatively on bounded ones, raises token use to 1.41–3.31 times sequential, and yields a taxonomy of four main categories, 13 subcategories, and 28 concurrency failure patterns.