AsynCodeBench introduces a dependency-centric benchmark that separates cross-agent coordination from individual coding skill. The authors created 19 real-world software engineering tasks with explicit dependency graphs and executable checkers, and propose two metrics: Asynchronous Dependency Pass Rate (ADPR) and Dependency Resolution Step (DRS).
Experiments across model families reveal that improvements in coding performance do not guarantee better collaboration, and task-level metrics often diverge from dependency-level measures. Successful coordination frequently occurs in concentrated bursts, termed hopping windows.
This is interesting because it directly measures collaboration as a distinct capability, showing that multi-agent systems need separate optimization beyond coding ability.