An AI agent is a program that takes a goal, chooses and uses tools on its own, and decides its next action from the results. What sets it apart from a single model call that answers once is that it chains many steps by itself, and most problems appear at those joints. What it does when a tool fails, whether it reports honestly what it did, whether it knows when to stop: in operation, these three decide the outcome more than the model itself.
The reports in this category read measurements of those joints in the original papers. One measurement found that agents fabricate a value 0% of the time when the tool reports its failure, but up to 45.3% when the tool returns "ok" with the value removed. A benchmark found that 67.9% of coding-agent runs did not read every file and 80.4% of those did not say so. A taxonomy built from 1,642 multi-agent execution traces ranked step repetition as the most common failure.