Self-improving AI makes most gains by rewriting its own prompts and tools — not its weights — because cheap iteration beats expensive retraining.

This reads the question as being about the split identified in Do self-improving agents really split into two distinct loops?: a slow parametric loop that rewrites the foundation model's weights, and a fast non-parametric loop that rewrites everything around the model — its prompts, its memory, its tools. Progress concentrates in the fast loop for a plain economic reason: scaffold edits are cheap and reversible, while weight updates are expensive, slow, and hard to undo. When improvement is a matter of trying, checking, and trying again, you want the cheapest possible iteration — and that's exactly what the non-parametric loop offers.

That matters more than it first appears because persistence, not raw quality, is what actually wins long-horizon agent work. What predicts success in ultra-long-horizon agent tasks? found across 17 frontier models that the dominant predictor of success was the willingness to run repeated edit-incorporate cycles within a budget, not how good the first attempt was. A fast, reversible loop is the substrate that makes persistence affordable — you can afford to keep going because each step is cheap. The slow loop can't sustain that cadence.

There are also structural reasons the fast loop is where the safe gains live. Can agents learn new skills without forgetting old ones? shows that storing skills in an external, composable library lets an agent keep learning without the catastrophic forgetting that weight updates invite — the knowledge accumulates outside the model rather than overwriting it. Should successful and failed episodes be processed differently? pushes further, treating successes as concrete demonstrations and failures as abstracted lessons, all held in context rather than in weights. The non-parametric loop isn't just cheaper; it sidesteps a failure mode that the parametric loop is prone to.

But the concentration is also a symptom of a limit, not only a virtue. Can models reliably improve themselves without external feedback? argues that purely self-directed improvement stalls on a generation-verification gap, diversity collapse, and reward hacking — and that the methods which actually work smuggle in external anchors like past model versions, judges, or tool feedback. The fast loop is exactly where you can plug those anchors in cheaply: user corrections, tool results, natural-language critiques. Can natural language feedback overcome numerical reward plateaus? shows that when numerical rewards plateau, a chain-of-thought critique — a non-parametric signal — can break through, because scalar rewards discard the 'why' that language carries (a point echoed in Can scalar rewards capture all the information in agent feedback?).

The surprise worth taking away: the fast loop's dominance may be a ceiling in disguise. Can AI systems improve their own learning strategies? points out that today's fast loops still run on fixed, human-designed metacognitive scaffolds that break under domain shift — the agent optimizes its prompts and memory but can't yet rewrite the strategy that decides how it optimizes. And Do stronger models always evolve harnesses better? finds that the ability to benefit from scaffold edits peaks at mid-tier models and falls off at the frontier, where strong models struggle to follow their own harness faithfully. So the fast loop is where progress concentrates precisely because it's cheap and reversible — but the same cheapness means the deeper, stickier gains still wait behind the slow loop nobody wants to pay for yet.