An Architecture for Long-Horizon Agents Built on Levels, Ticks, and Cascaded Intelligence

Summary

Language-model agents are increasingly expected to handle work lasting days or weeks, but such tasks outlast individual context windows, processes, and periods of human attention. This paper argues that an agent must first run continually without forgetting before it can learn continually, and that the required substrate belongs primarily in the harness around the model. It identifies seven bottlenecks in long-horizon operation and proposes a three-part hierarchical architecture. Levels are indexed by time scale, with each maintaining a bounded file that summarizes the level below; a clocked tick defines the unit of autonomous action; and cascaded intelligence escalates work to a more capable model only after a review failure. In a ten-day campaign, an agent using the architecture reproduced a published reinforcement-learning result with a person attending once per day. The agent retained the thread across every context reset and session boundary, while knowledge written early in the campaign changed later behavior without modifying model weights. The paper also indicates where learned components could be introduced, and concludes that continual learning requires a substrate that outlives each context and process, with the harness’s existing checks providing a natural location for learning.