Language‑model agents are increasingly tasked with work that spans days or weeks, such as operations remediation or a research programme. Such tasks exceed any context window, any single process, and the interval a human can continuously attend. We argue that a long‑horizon agent must run continuously without forgetting in order to learn continually. This capability resides in the harness surrounding the model rather than in the model itself. From the long‑horizon setting we derive seven bottlenecks and address them with a three‑part hierarchical architecture: (i) levels indexed by time scale, each keeping a bounded file that summarises the level below; (ii) a clocked tick that serves as the unit of autonomous action; (iii) cascaded intelligence, where work is escalated to a more capable model only after a failed review. In a ten‑day campaign we built an agent on this architecture that reproduced a published reinforcement‑learning result with a human attending only once per day. The campaign demonstrated that (1) the agent preserved the task thread across every context reset and session boundary; (2) operating knowledge written early altered later behaviour without changing model weights; (3) we identified where learned components could be inserted into such a system. Overall, our experience suggests that continual learning for long‑horizon agents requires a substrate that outlives every context and process, and the checks already performed by the harness are the natural place for a learner.
Review: The proposed architecture combines hierarchical summarisation, tick‑driven autonomy, and conditional escalation, offering a scalable blueprint for long‑term autonomous agents and showing that continual learning is feasible with minimal human oversight.