Can external state caches let models solve harder problems?

Prime Agent starts from a limit on the model itself: an LLM is "a bounded sequential processor whose next decision can use only state information exposed in its weights and active context." Long-horizon agency needs more state and computation than that, so the harness supplies a substrate. The introduction imagines that substrate as a state information cache with four levels. Model weights are L0 and active context is L1. A persistent REPL plus recursive subagents form L2, and disk-backed history, memories, and skills form L3. The paper calls the result "more von Neumann-like," because the model can read, transform, and write addressable state outside the instruction it is currently generating. The abstract reports that ARC-AGI-3 RHAE Best@1 rises from 30% to 95.5%, and that the harness "matches or exceeds native and popular harnesses" on long-context coding, GPU-kernel generation, emulator construction, and autonomous nanoGPT speedruns.

The mechanism is a division of labor. Prime Agent "standardizes execution, recovery, verification, and resource accounting while leaving strategy construction to the model." The authors call this a "low-friction, expressive membrane" that keeps harness failures from turning into model failures and moves measurement toward "the model's true maximal underlying capability." Persistence comes from Continual Harness, which carries histories, memories, skills, prompts, and subagent specifications across trajectories. Recursive subagents coordinate by direct agent-to-agent communication, and an Agents View lets humans inspect and manage daemon-backed sessions. The introduction treats agentic compaction, in which a model prunes its own context, as the first form of context management. It presents the other levels as what became necessary once the full information state outgrew weights and tokens.

Against the nearest notes, this adds two things. Where does agent reliability actually come from? already argues that the harness hosts externalized memory, skills, and protocols. Prime Agent orders that state by how far it sits from the weights, and it adds a measurement argument: a badly built harness contaminates what a benchmark says about the model. Can an external manager handle context for frozen agents? works at L1, learning what to keep in context. Prime Agent moves state out of L1 into levels the model addresses itself. Can delegation teach models to manage context more actively? reads subagents as a context tool, which matches their placement at L2 here. Should we evaluate deployed agents as whole environments instead? points the same way on evaluation, and the Agents View is one concrete human-facing surface.

The excerpt is silent on what produced the gain. It does not name the baseline harness or the models used, give sample sizes, define RHAE Best@1, or ablate the four levels, so the 30% to 95.5% jump cannot be credited to any one component. The cache mapping is offered as an image, and the excerpt gives no test of it. The discussion also undercuts its own measurement claim. It concedes that "models still experience friction" when allocating subagents, managing retained information, and refining reusable state, and that many harness capabilities go underused "because current models were not trained to operate them." The expectation that model-harness co-learning becomes "the dominant route to new long-horizon capabilities" is a forecast, not a result. At the strength the evidence allows, a score from this harness measures a model paired with a harness the model has not been trained to use fully, and it is not a clean reading of maximal capability.