What's this about?
A harness aimed at "long-horizon" autonomy — pushing agents toward tasks that span a full day's worth of work rather than a single bounded prompt (12↑). It's one of a visible surge of long-horizon harness papers this week (alongside LongHorizon-Harness and MerchantBench).
What exactly is a "long-horizon" task?
A task with many steps stretched over a long time, where the agent must maintain a goal, track state, and adapt as it goes — running a project, not answering a question. It's the hard frontier for agents, which tend to be strong in short bursts and drift over long runs.
Why does it matter?
The week keeps circling the same weakness — agents degrade over long horizons — from multiple angles (long-context degradation, PAST-Bench, MerchantBench). A dedicated harness for day-long tasks is the engineering response to the benchmarks that keep exposing the gap.