When Seeing Is Not Knowing: Claim-Bounded Observability for Agentic Systems

Passive observation is attractive for agentic systems because it can discover tool activity without instrumenting every endpoint. The danger is evidentiary overreach: observing protocol traffic is not the same as establishing which agent caused it, whether a tool actually executed, or whether an external effect occurred. We propose claim-bounded observability, an AgenticOS abstraction in which each finding states both the claim supported by the available evidence and the nearest stronger claim that remains unsup- ported. Using the Model Context Protocol (MCP) as a motivating case, we define an evidence ceiling over observation and vantage, a compact claim lattice from protocol presence to external effect, and an overclaim-rate metric for evaluation. The position is deliberately conservative: abstention is a correct systems outcome when the observation source cannot support a stronger inference.