The paper provides the sample, protocol and failure analysis but no public runnable artifacts; results are not independently reproduced.
FeatureArc finds coding agents inherit defects across feature stages
2026-10-27Frontier watchWATCH
FeatureArc evaluates coding agents on features that evolve through two to ten dependent stages in one persistent workspace. Across ten agents and 73 manually reviewed tasks, pooled success fell from 62.7% on the first stage to 28.5% on the completed feature. The authors report that 72.9% of later failures inherited an earlier defect, but publish no runnable harness or code.
What this means for you
For teams running long coding tasks, stage-one success is a weak proxy for feature completion. The result supports adding cumulative regression gates between stages, while the missing artifacts keep it from being a deployable benchmark.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Sample | 10 coding agents; 73 manually reviewed tasks; 2-10 stages; four programming languages |
|---|---|
| Protocol | one persistent session and workspace; no external test feedback or rollback |
| First Stage Success | 62.7% pooled |
| Whole Feature Success | 28.5% pooled |
| Inherited Defects | 72.9% of failed later stages |
| Artifacts | no public harness or code identified |
| Time Basis | arXiv submission 2026-10-27T15:45:50Z; inside overlap and 13m25s before the prior checkpoint |
Technical impact 3
Engineering value 4
Adoption signal 1
Maturity 1
Verification cost 3
Risk 2