A Cross-Level Framework for Human–Agent Organizations: Organizational Characterization Supporting Governance, Security, and Compliance Assessments

AI agents increasingly perform organizational activities, contribute to substantive decisions, coordinate work, and participate in governance. Existing research examines human–AI collaboration, agent capabilities, governance, organizational structures, and technical trust boundaries at different levels. This paper connects these perspectives through an assessment-oriented framework for Human–Agent Organizations. It distinguishes four organization-level components: Organizational Actor Composition (COMP), Agentic Share (ASHR), Governance Mode (GOVR), and Organizational Autonomy (OAUT). Together, they form a compact organizational profile, 𝒪 = ⟨COMP, ASHR, GOVR, OAUT⟩, supported by more detailed evidence about actors, governance, dependencies, and trust boundaries. A documented Measurement Protocol (ℳ) specifies which organizational activities, substantive decisions, and governance functions are assessed, how assessment units are classified and weighted, and how human and agent contributions are coded and aggregated. It also defines how uncertainty and protocol changes are documented. Agentic Share summarizes agent participation under declared scoring rules, rather than providing a calibrated interval-scale measure. The framework keeps individual actors' capability, authority, autonomy, and operational responsibility distinct from institutional accountability. It also records the external trust-domain status of persistent third-party agents. The resulting organizational profile can inform subsequent governance, security, and compliance assessments, but does not determine their outcomes. An exploratory multi-model coding study applied frozen, case-specific instruments to 16 synthetic cases using four separately instantiated AI coders, yielding 64 selected coder-case responses. The selected coder responses agree unanimously on 90 of 96 case-variable cells across six structural inputs. Structural disagreements concerned organizational composition, governance aggregation, and continuity. Qualitative inspection additionally showed interpretation sensitivity in activity classification and Agentic Share. These descriptive findings apply only to the controlled synthetic cases. They do not establish human-expert or psychometric reliability, population-level performance, construct or external validity, comparative advantage, or practical utility. The study therefore examines feasibility and interpretation boundaries rather than validating the framework.