The SKILL.md Fallacy: Coding Agents Should Use Ephemeral Subagents

rseroter · x · 2026-08-18

Ali Afshar critiques the industry obsession with SKILL.md files for coding agents. The article argues that while progressive disclosure saves tokens for small tasks, flat skill libraries suffer catastrophic non-linear failure as complexity grows. The correct system primitive is ephemeral subagent delegation with isolated process boundaries, not runtime prompt injection.

Original post →

More from coding & agent

coding & agent channel →