Long-horizon is a term used in AI for tasks that must be carried out over many steps and an extended period (long-horizon tasks) and for the reasoning such tasks require (long-horizon reasoning). Unlike a single question and answer, these tasks require breaking a goal into sub-steps, keeping track of intermediate results and correcting errors along the way until completion. The expression originally referred to reinforcement learning problems in which rewards arrive far in the future.
As AI agents take on lengthy work such as software development, computer operation and research, long-horizon capability has become a key measure of model performance. Because small errors accumulate as the number of steps grows, the length of task a model can complete autonomously is sometimes used as an evaluation metric.