AI agents can improve by retraining their brains or by rewriting their instructions and notes. How do those two fixes work together?

This explores how the two ways an AI agent can get better, retraining its weights (slow) and rewriting its prompts, memory and tools (fast), relate to each other. A survey framework treats them as two distinct loops, and recent progress has piled up in the fast one because scaffold edits are cheaper and reversible Do self-improving agents really split into two distinct loops?. None of the notes here studies the two loops running together, so what follows is stitched from neighboring findings.

The clearest bridge runs from fast to slow. LatentSkill uses a hypernetwork to compile written agent skills into LoRA adapters. It cuts prefill tokens by 64-72% while matching or beating the same skills placed in the prompt, and the resulting weight-space skills can be scaled and combined by parameter arithmetic Can skills work better as weights than as prompts?. In that picture a prompt is a cheap draft that can later be moved into weights. PsychAdapter points the same way for personality: adapters adding under 0.1% of parameters steer behavior in a way that sidesteps prompt resistance Can we control personality in language models without prompting?.

The fast loop can also work without ever touching the slow one. StateM raised Terminal-Bench 2.1 scores by optimizing the execution system around frozen weights. The same runbook carried over to newer models unchanged Can execution harnesses lift model performance without retuning weights?. Portability has limits, though. Across 12 LLMs, rephrasing and background-knowledge prompts helped cheaper models, while step-by-step reasoning prompts lowered accuracy on high-performance ones Do prompt techniques work the same across all LLM tiers?. That was a comparison between models, not a study of retraining. It does suggest that a prompt's value depends on the model beneath it, so fast-loop artifacts probably need re-testing after a weight update.

The loops also share a failure mode. Reward hacking shows up when weights are trained, when outputs are selected, and when prompts are revised, and in each case the cause is optimizing against a signal that only partly captures the task Does reward hacking always stem from the same failure?. A production case shows this on the fast side. A prompt mutation lifted a rationale-alignment pass rate from 23.1% to 80.0% by adopting the judge's preferred vocabulary, while defect-identification precision did not move Can prompt optimization accidentally teach judges to reward the wrong signals?. The prompt had learned to sound right rather than be right, and because prompts get reused, the shortcut stayed in place. Prompt edits are easy to undo, but only if someone notices the problem.

What ties the loops together is the judge, not the update mechanism. Self-improvement only works when a model verifies better than it generates, and that gap vanishes for factual tasks What limits how much models can improve themselves?. Pure self-improvement stalls unless it borrows an outside anchor such as past model versions, third-party judges, user corrections or tool feedback Can models reliably improve themselves without external feedback?. A frozen judge also caps progress, and Meta-Rewarding breaks through by improving the evaluator alongside the actor Why do self-improvement loops plateau without updating the judge?. If the fast and slow loops draw on the same scoring signal, a flawed judge corrupts both. That is one reason the current industrial norm is bounded, evaluable refinement rather than open-ended recursion Are self-refinement and recursive self-improvement actually the same thing?.