A controlled study isolates coding-agent harness effects
A preprint and MIT-licensed Yuj harness compare the same coding model under a full chronological context against a treatment that compresses older tool output and intervenes on repeated or stalled work. Under a 20,480-token window and fixed 480-second endpoint on 169 SWE-bench Verified tasks, complete solutions rose from 43 to 72; the frozen treatment was also tested with three additional model designs.
Why it matters: Coding-agent scores can move materially without changing model weights. Model selection, procurement, and safety evaluation therefore need to name the harness, context policy, tools, recovery rules, runtime limits, and verifier as part of the tested system.