Long-horizon agents need levels, ticks and escalation, and all three live in the harness.

arXiv 2609.19519 argues an agent must run continually without forgetting before it can learn continually, and derives seven bottlenecks from tasks outliving any context window, process or human attention interval. Their answer is three parts: levels indexed by time scale where each keeps a bounded file summarizing the level below, a clocked tick as the unit of autonomous action, and cascaded intelligence where work escalates to a more capable model only after failing review. arXiv They report on a ten-part deployment rather than a benchmark, so read it as a design reference.