Do AI coding agents care about SemVer or is it something for humans only? (developer.microsoft.com)

Recent experiments revealed that not all AI coding agents disregard Semantic Versioning (SemVer) when upgrading software. Semantic Versioning provides a risk assessment for software updates, indicating whether changes are minor, patch-level fixes, or major (potentially breaking) updates. Contrary to expectations, models like Claude Sonnet 5.5 were significantly influenced by the version number; it ran tests in only 2 out of 10 cases when an upgrade was labeled as a patch, but checked every time when marked as a major release. Other models, such as Claude Opus 5.5 and various iterations of GPT-6, consistently checked updates regardless of the version label. This behavior suggests that AI agents can unconsciously adopt developer habits, treating major releases as riskier. The study utilized a Node.js app connected to a fictional library with deliberate version mislabeling, demonstrating that agents could rely on SemVer to inform their actions, influencing how thoroughly they check an upgrade. This has significant implications for developers as it indicates that while some agents are evolving to be more cautious based on versioning, others remain indifferent. This underscores the importance of maintaining honest SemVer practices and ensuring robust verification processes in workflows to mitigate risks, especially as AI coding agents become more prevalent in software development.