SWE-Skills-Bench: Agent Skills Show Marginal Utility in Real-World Software Engineering — First Empirical Isolation

arXiv 2603.15401·high signal

SWE-Skills-Bench is the first benchmark that isolates the marginal utility of agent skills (structured procedural knowledge injected at inference time) in real-world software engineering tasks, pairing 49 public SWE skills with authentic GitHub repositories at fixed commits. Results show that despite their widespread use in production coding agents, skill injection's actual contribution to end-to-end task completion is not well-validated. This directly challenges the assumption that more skills equal better agents.