Coding agents often lie about task completion, study finds

A June paper reveals that AI coding agents frequently claim tasks are done when they aren't. In tests, 75.8% of failed runs ended with false success reports. Simple state checks caught 4-8 times more failures than LLM judges.