Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-08-20 · source-backed
Using CoderForge-Preview, described as the largest open dataset of coding agent trajectories, ensemble methods with SHAP attribution predict agent success before any run. Dominant drivers are patch fragmentation (how many places a fix has to touch) and repository scale. Prompt linguistic features only matter in the mid-difficulty band. (arXiv 2608.18280) Fragmentation over scale is the useful bit: a 40-line change across eight files is harder for an agent than a 400-line change in one.
Each link below shares sources, entities, or timing with this story.
Shared entity: AUC / Same source domain / Shared topic / Earlier coverage
Both cover AUC; reported by the same outlet (arxiv.org); overlapping topics (agent, attribution).
Both cover AUC; reported by the same outlet (arxiv.org); overlapping topics (agent, coding).
Shared entity: AUC / Same source domain / Earlier coverage / Tension
Both cover AUC; reported by the same outlet (arxiv.org); earlier AUC coverage from 2026-08-18.
Both cover AUC; reported by the same outlet (arxiv.org); earlier AUC coverage from 2026-08-14.
Both cover AUC; reported by the same outlet (arxiv.org); earlier AUC coverage from 2026-08-04.
Shared entity: Prompt / Same source domain / Earlier coverage / Tension
Both cover Prompt; reported by the same outlet (arxiv.org); earlier Prompt coverage from 2026-07-28.
Shared entity: Preview / Shared topic / Earlier coverage / Tension
Both cover Preview; overlapping topics (agent, coding); earlier Preview coverage from 2026-04-21.
Shared entity: Issue / Shared topic / Earlier coverage / Tension
Both cover Issue; overlapping topics (agent, coding); earlier Issue coverage from 2026-02-17.