What Counts as the Agent? Boundaries and Measurement in Agentic Systems

In current agentic systems, the term agent is often tied to implementation choices. Yet measurements of system properties such as autonomy, control or conformity depend on which parts of a system count as the agent. We argue that implementation structure and a weak perception/action criterion constrain possible agent boundaries without necessarily selecting one. We therefore propose evaluating candidate boundaries relative to a fixed property: an adequate representation must preserve the execution and structural distinctions needed to measure that property. A minimal example shows that plausible boundaries can disagree about the measurement of a specific property. We outline how system evidence (such as traces) can constrain and assess these candidates. Even complete evidence need not settle which bearer an unqualified agent-level claim refers to. When the remaining adequate candidates disagree, we call the claim boundary-undetermined; when they agree, the measurement does not require a unique boundary.