arXiv cs.AIPaper
A Case Study in Assuring AI-Written Software
This is the reckoning: agents can outpace human code review, and the usual safeguards don't prevent failures at scale. The operator learned that tests break silently, monitors measure proxies instead of outcomes, and repair agents can cause cascading outages. If you're using agents on production systems, this is a hard-earned field report on where your safety net will tear.