Long-horizon agents need a harness that outlives every context
Sep 17 Nijkamp et al.: levels + clocked ticks + cascaded intelligence — a ten-day campaign with once-a-day human attention kept thread across every reset; continual learning belongs in harness checks.
Source: An Architecture for Long-Horizon Agents: Levels, Ticks and Cascaded Intelligence
Nijkamp, Koul, Pakhomov, and Pang start from the factory premise factories keep rediscovering: a multi-day task outlives any context window, any process, and any interval a human can attend. Before an agent can learn continually, it must run continually without forgetting — and that ability lives in the harness around the model, not in weights. They derive seven bottlenecks from the long-horizon setting and answer with three architectural parts: (i) levels indexed by time scale, each keeping a bounded file that summarises the level below; (ii) a clocked tick as the unit of autonomous action; (iii) cascaded intelligence, escalating to a more capable model only after failing review.
The empirical receipt is a ten-day campaign that reproduced a published reinforcement-learning result with a human attending once a day. Across every context reset and session boundary the agent kept the thread; operating knowledge written early changed later behaviour with no weight updates; and the authors locate where learned components would enter such a system — inside the checks the harness already runs. Continual learning, on this view, needs a substrate that outlives every context and process; the review gates are the learner’s home, not an afterthought prompt.
Read next to Anthropic’s effective harnesses for long-running agents (initializer + incremental clean commits) and Fan et al. on which harness components move. Those tune a session-spanning wrapper; this paper asks what architecture keeps a week-long factory coherent when the human is mostly gone.
Why it matters for code review
Steady overnight momentum is the short form of the same claim: if state and review live only inside a chat context, continuous factories forget between resets and FOMO for “more model” cannot recover the thread. Treat levels, ticks, and cascaded review as reviewable systems — bounded summaries, clocked autonomous units, escalate-only-after-fail — so multi-day verified throughput compounds past peak-session heroics instead of restarting from vibes every morning.