Beyond Direct Access: Reachable Geometry and Human Agency Execution Boundaries for Governing Transitive Agent Capabilities

Agentic AI systems are commonly constrained through sandbox boundaries, tool permissions, network restrictions, identity controls, and interface-level access policies. These mechanisms describe capabilities intentionally exposed to an agent, but they do not necessarily describe the full set of effects the agent can produce through intermediary systems. This paper introduces Reachable Geometry, a runtime representation of state-dependent executable effect paths, and integrates it with the Human Agency Execution Boundary (HAEB) for authority-sensitive execution decisions. A temporally constrained worked analysis of the 2026 OpenAI–Hugging Face incident — extended through the incident's full disclosed timeline, including its eventual cross-organizational escalation into a third party's production infrastructure — motivates three distinctions: direct access is not equivalent to executable reach; reachability expansion does not imply authority expansion; and repair does not imply reauthorization. The cross-organizational escalation further surfaces an open observability problem for the framework itself — HAEB's authority evaluation presupposes that the governing party can sufficiently observe or reconstruct the relevant executable-effect graph, an assumption this case suggests may not hold once effective reach crosses an organizational boundary — discussed here as a limitation and open research question rather than a resolved contribution.