Agent datasets need failure and recovery traces

NVIDIA and Hugging Face argue that useful agent data should include tool failures, retrieval problems, workflow recovery, safety behavior, and multi-step traces rather than only successful task outputs. Why it matters: Agent quality depends on what systems learn and evaluate when tools break, context is incomplete, or a workflow needs recovery. Datasets that preserve these traces make reliability work more concrete.

Try this: When building an agent eval set, save at least one example each for tool failure, bad retrieval, recovery after an error, and a case where the agent should stop or ask for review.