New paper proposes hierarchical architecture for long‑horizon AI agents

To validate the design, they ran a ten‑day campaign in which the agent reproduced a published reinforcement‑learning result while a human intervened only once per day. The experiment showed the agent preserved the task thread across every context reset, allowed knowledge written early in the run to alter later behavior without any change to model weights, and illustrated where learned components could be integrated into such a system. The authors conclude that continual learning for agents depends on a substrate that survives all context and process boundaries, and that the existing harness checks are the natural place for a learner.