Self-improving agents get a hype cycle in one afternoon

The Astra post ("can read and write weights to improve Qwen") was followed almost immediately by "The 'Astra' post is misleading." Commenters on the rebuttal were unimpressed. One said anyone who read past the title already knew, another said the original explained it better. Nvidia's open-source auto-ontology drew a sharper jab: "Fixing unreliable agents by adding more unreliable agents." Others asked what an ontology gives you that a memory system plus skills for your APIs and CLIs doesn't. "Ten harness patterns that you can apply to a real repo" sits in the same cluster.

The pattern here: the interesting work has moved from the model to the scaffolding around it (harnesses, ontologies, memory, weight edits). Readers now treat every new scaffold as unproven until someone shows it beating a boring baseline. Titles overclaim, the body is fine, and the community fact-checks in real time.

The counterpoint is that nobody is dismissing the direction. The pushback is about packaging and claims, not about whether agents that modify themselves or their context are worth building.