The same language model can behave very differently under distinct harnesses. This study keeps the model and the task fixed while comparing two harness configurations across three coding benchmarks. The control configuration feeds the full conversation in chronological order; the treatment configuration shortens older tool results once the context window fills, freeing space and reacting to repeated or stalled work.
Under a tight context window the treatment raises the mean per‑task fail‑to‑pass fraction (F2PF) in all three pressure settings and increases the number of complete solutions on SWE‑bench Verified and SWE‑bench Pro. The tight‑window Verified experiment uses 169 tasks, a 20,480‑token window and a fixed 480‑second attempt endpoint; in this cohort the treatment lifts average F2PF from 28% to 49% and complete‑solution rate from 43% to 72%. Without model‑specific retuning, the same frozen treatment also improves both endpoints for three additional models with different architectures.
In the wide‑window Qwen3.6 comparisons, overall outcomes on Verified and Pro are similar, while FeatureBench retains a higher mean per‑task F2PF under treatment. In the wide‑window Verified cohort the treatment consumes fewer prompt tokens per turn. Because changing the harness alters what unchanged model weights can achieve, coding‑agent evaluations should treat the model and harness together as the tested solver.
Blogger's Review: The work highlights that the design of the execution environment is as crucial as the model itself; even with frozen weights, subtle harness choices can dramatically sway performance.