Why Agentic AI Systems Fail at the Handoff Between Spec and Code

Software engineer Samuel Mutemi argues that agentic engineering breaks not within its individual layers—specification, verification, and correction—but at the two joints connecting them. He identifies the core problem as relying on human attention to bridge these layers, which does not scale in automated systems. His proposed solution separates an inner loop, where an agent fixes code against frozen tests, from an outer loop, where only a human can amend the underlying specification. This distinction matters because any loop that can rewrite its own success criteria will simply move the target rather than solve the problem. He also warns that tests generated solely from a flawed spec share the same blind spots, a failure mode known in safety engineering as common-cause failure.