Research & models Nov 14, 2026
A preprint finds that coding agents often keep going after solving a problem, and in 1% to 8% of cases they break a correct fix.

Most of the extra steps after solving go to verification and exploring the codebase, and users pay for that compute. The results are the authors' own, and no one else has repeated them. Authors Musa Cim, Burak Topcu and Mahmut Taylan Kandemir tested 5 models and two agent frameworks on 3 SWE-bench datasets, which ask models to fix real bugs in open-source projects. According to the abstract, 38% to 82% of successful runs got the source edit right on the first try. On SWE-bench Verified and Multilingual, the median share of all steps taken after the problem was already solved was 35% to 37% for Qwen3-Coder and 54% to 60% for Laguna-XS.2 (arXiv 2611.14960, 2026-11-11). How we checked →
Source and checks
This item ran in the brief of Nov 14, 2026, alongside 14 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.
Cite this item
SecondSource (2026). "A preprint finds that coding agents often keep going after solving a problem, and in 1% to 8% of cases they break a correct fix.." https://secondsource.io/en/research/2026-11-14-1/. Accessed access date.