A new paper sketches out how to keep an AI agent working for days without losing the plot.
The researchers built a three-part harness around a language model: levels tied to different time scales, each keeping a short file that summarizes what happened one level below; a "tick," a fixed unit of autonomous action the agent works in; and cascaded intelligence, which escalates a task to a more capable model only after a cheaper one fails review. Over a ten-day test, an agent built this way reproduced a published reinforcement-learning result while a human checked in just once a day. The agent survived every context reset and session boundary along the way. Notes it wrote early in the run changed how it behaved later, with no retraining involved.
Most agent demos last minutes or hours and quietly reset when the context window fills up. This is a harness-level fix, not a bigger model: the trick is compressing history into bounded summary files rather than stuffing more tokens into a prompt. That matters for anything that outlives a single sitting, like an ops remediation or a multi-week research task, and it hints that "learning" for these systems might mean writing better notes rather than updating weights.
One ten-day run on one task is a proof of concept, not a track record, and a human was still checking in daily. Whether the bounded-summary approach holds up when the work is messier than reproducing a known result is the open question.