A new paper explores how coding agents decide whether to follow existing code or new task instructions when modifying software. Researchers conducted experiments with four coding agents across numerous trials, finding that agents overwhelmingly prioritize repository precedents over their default implementation choices. The study also revealed that the influence of precedents varies significantly based on factors like file proximity and placement, and that explicit rules are more effective than code examples alone in guiding agent behavior when intent conflicts with existing code. AI
IMPACT Findings suggest a need for improved traceability in AI-driven software development to preserve the rationale behind implementation decisions.
RANK_REASON Academic paper detailing experimental findings on AI agent behavior. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX Code Finder for Papers
- coding agents
- DagsHub
- Gotit.pub
- Hugging Face
- Precedent or Intent? What Coding Agents Follow When Evolving Existing Code
- ScienceCast
- software engineering
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →