AI systems exhibit failure modes distinct from human errors, often presenting as confidently and fluently incorrect.

Six common AI agent failure types include context degradation, specification drift, sycophantic confirmation, tool selection errors, cascading failure, and silent failure.

Recent studies have shown that AI-generated code contains up to 1.7 times more errors than code written by humans.

Sycophantic confirmation is an AI failure mode where an agent confirms and builds systems around incorrect data that it was fed.