AI researchers push reliability tests for agent systems (opens in new tab)

A group of AI research articles and preprints focused on the same operational problem: large language models and agentic systems are being used in longer, tool-based workflows while evaluation methods still miss important failures. Nature argued that large language models require “capability-based monitoring” as a new form of oversight, while Oxford Martin School research said LLMs can hallucinate with certainty despite knowing the correct answer. Microsoft Research separately clarified that ...