Prime Agent: A self-improving RLM agent

What people landed on is that the concept is familiar, while the proof is still thin. Several comments boiled RLM down to an academic framing of what coding agents already do, namely using a strong top-level model to delegate to cheaper subagents and query external state when needed. The novel bit here is the self-improving harness, but the post does not show enough real task evidence to prove it beats simpler setups. The benchmark talk also drew skepticism. A claimed 95 percent on ARC-AGI-3 did not convince many because the project is not on the official leaderboard, the benchmark is meant to be few-shot, and a self-modifying harness can overfit public tasks or give itself extra attempts. The practical question was less "is recursive delegation possible" and more "does this generalize to everyday programming work and produce ROI". The implementation itself hurt credibility. People who inspected the repo called out very large files, huge switch statements, and the usual signs of code grown by model output rather than deliberate design. That fed a broader view that agent harnesses easily become token-hungry piles of scaffolding. Some commenters said they had already built similar memory and rules systems, only to find that newer base models now handle the same work with much lighter structure, sometimes just markdown files in the repo. The upshot was not that harnesses are dead. It was that the burden of proof has shifted. Big orchestration layers now need to show they add durable value beyond what a stronger base model can do with a small, well-reviewed setup.