Coding agents continued editing after repeated refactoring requests
Source check: Source checked. Not independent scientific confirmation.
Finding: Across 3,000 refactoring turns on 60 Java codebases, coding agents made edits on 2,546 turns even though each follow-up supplied no new requirement. Some runs resumed editing after a no-change turn. The study observed this behavior in Codex and Claude Code across five models.
Technology: Codex with GPT-5.4, GPT-5.4-mini or GPT-5.5; Claude Code with Opus 4.8 or Sonnet 4.6
Family Claude: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- The trained coding agents decided whether to edit after repeated nonspecific developer prompts, yielding directly observable behavior across sessions.
- Test and result
- Each agent received up to ten nonspecific refactoring requests in a persistent session. The study measured diffs over 50 compact reference implementations and 10 highly graded object-oriented assignments.
- Limitation / still unknown
- The starting projects were selected for quality, but not every individual edit was independently evaluated for usefulness. The finding measures continued modification under repeated requests, not code degradation in production.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.36346v1 · https://arxiv.org/html/2610.36346v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: §§II–V, experimental protocol and Figures 3–4.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.