An Architecture for Long-Horizon Agents: Levels, Ticks and Cascaded Intelligence
Erik Nijkamp, Anurag Koul, Egor Pakhomov and Bo Pang at Salesforce AI Research argue that a long-horizon agent must run continually without forgetting before it can learn continually, and place that capability in the harness rather than the model.
Ask this paper
Seven bottlenecks, three abstractions. Levels indexed by time scale each keep a bounded file summarising the level below; a clocked tick is the unit of autonomous action; cascaded intelligence escalates work to a more capable model only after it fails review.
A ten-day autonomous campaign. An agent built on the architecture reproduced a published reinforcement-learning result with a human attending once per day.
The thread survived every reset. The agent kept task state across all context resets and session boundaries during the campaign, which is the property the level-and-tick structure is designed to supply.
Written knowledge changed later behaviour. Operating notes recorded early in the campaign altered how the agent acted later with no change to model weights, which is the paper's argument for a substrate before a learner.
It names where a learner would attach. The checks the harness already runs are proposed as the natural insertion point for learned components.
Abstract
Language-model agents are increasingly asked to carry out work spanning days or weeks, such as an operations remediation or a research programme. Such a task outlives any context window, any process and any interval at which a person can attend. In this paper, we argue that a long-horizon agent must run continually without forgetting before it can learn continually. This ability lies in the harness around the model rather than in the model itself. We derive seven bottlenecks from the long-horizon setting and answer them with a hierarchical architecture of three parts: (i) levels indexed by time scale, each keeping a bounded file summarising the level below; (ii) a clocked tick as the unit of autonomous action; and (iii) cascaded intelligence, where work is escalated to a more capable model only after failing review. We report on a ten-day campaign in which an agent built on this architecture reproduced a published reinforcement-learning result with a human attending once a day, and show (1) the agent kept the thread across every context reset and session boundary of the campaign, (2) operating knowledge written early changed later behaviour with no change to model weights, and (3) where learned components would enter such a system. Overall, our experience suggests continual learning for these agents needs a substrate outliving every context and process, and the checks the harness already runs are where a learner belongs.