Does recursive self-improvement start with harness engineering?

Weng argues the recursive self-improvement (RSI) loop in modern AI need not take the form Yudkowsky described — "an AI uses its current intelligence to improve the cognitive machinery that produces its intelligence" — as a model rewriting its own weights. The feedback loop, she writes, "may indicate the model rewriting its own weights directly, or more broadly the model improves the training pipeline and the deployment system," and she treats the deployment layer, the harness, as "as important as the model's raw intelligence." Her prediction is explicit: "the near-term path of RSI is unlikely to start as a model directly rewriting its weights" — it instead runs through harness engineering, the system that "orchestrates execution and decides how the model thinks and plans, calls tools and acts, perceives and manages context, stores artifacts, and evaluates results."

The mechanism she proposes is a staged progression of what gets optimized: "instruction prompts → structured context → workflow → harness code → optimizer code." As models grow more capable, the target of improvement climbs this ladder toward more generic, less heuristic mechanisms — "the harness system itself becomes an optimization target, with fewer heuristic rules and more general mechanisms." She draws an explicit analogy to operating systems (a harness should "encapsulate complicated logic while keeping the interface simple") and to the history of prompt engineering: "manual prompt tricks became less central as instruction tuning and model reasoning improved, but the need to specify goals, constraints, context, and evaluation did not disappear." By that analogy she expects harness-level gains eventually to be "internalized into core model behavior," while the interface to external context and tools persists.

This reframes several library notes as evidence of rungs on the progression Weng describes, rather than as isolated results. Can language models build and maintain their own agent harnesses? supplies the empirical basis for her "deployment system" framing: HarnessDev shows the same model scoring differently under different harnesses, which is exactly the separation her argument needs. Does self-editing through reviewed commits improve agent performance? is a concrete instance of her "harness code" rung, with human review standing in for the heuristic rules she expects eventually to fall away. Can a routing harness generate its own training data automatically? makes a parallel bet that RSI's near-term engine is the deployed harness layer rather than direct weight modification. Can harness modules improve separately from benchmark data? targets the generalization problem her progression implies: a harness optimized on one benchmark is stuck on a lower rung unless its updates are built to transfer.

The excerpt is a position essay, not a measurement: Weng cites other groups' benchmark results (AHE on Terminal-Bench-2, Self-Harness on Terminal-Bench-2) in passing but reports no data of her own, and her claim that harness gains will eventually be "internalized into core model behavior" is a stated prediction, not something the post demonstrates. It also does not specify what would count as a harness update crossing into weight-level RSI, leaving the boundary between "the deployment system improves" and "the model improves itself" to the reader's judgment. The implication, held at that strength, is that progress toward RSI is for now better tracked by which rung of the stack — prompts, context, workflow, harness code, optimizer code — a given result moves, than by whether any single benchmark score went up.