The original title is: "Prose in the Control Plane: Why AI Agent Frameworks Are Not Engineering (Yet)"

Short summary

AI agent skill frameworks (Superpowers, Matt Pocock's skills, Agent Skills) rely on Markdown prose as their behavioral control layer, but prose is probabilistically interpreted by LLMs—unlike deterministic config files or type systems. The article identifies three failure modes: context drift over long sessions, scope expansion beyond user intent, and self-verification loops where the verifier suffers the same interpretation drift as the actor. Anthropic's own SWE-bench docs acknowledge performance varies with this scaffolding.

• •AI agent frameworks use prose Markdown as control plane, but LLMs interpret instructions probabilistically, not deterministically
• •Three failure modes: context drift over long sessions, scope expansion without confirmation, and verifier drift undermining self-checking
• •Current frameworks lack mechanisms to detect when agent interpretation has drifted from original intent

Generated with AI, which can make mistakes.