The Neutral Boundary

Autonomous agents are being composed into systems faster than anyone can account for what passes between them. The prevailing instinct is to make each agent more trustworthy: better models, better benchmarks, better guardrails. This paper argues that this instinct is aimed at the wrong place. In composed systems, we argue, the dominant failure mode is not a bad agent, but a bad handoff. A bad handoff is a work product or package that is structurally plausible, silently wrong, and consumed downstream as if it were evidence. We propose a theory of the neutral boundary: the claim that reliability in agentic systems must be established at the crossings where work changes hands, by a check that is neutral in a precise sense - its outcome owes nothing to any model, any domain, or any party to the handoff. We define neutrality in this agentic AI context, state the properties any neutral boundary should exhibit, show that the same boundary recurs at three levels - between stages, between workflows, and between organizations - and argue that recent independent work organizing the agentic-design field around handoff validity is evidence the field is converging on the same conclusion. The theory is falsifiable and we state how.