A Formal Architecture for Governed Enterprise AI Agents: Forms, Lifecycle, and Accountability

Enterprise AI agents are increasingly deployed across workflows, platforms, and organizational boundaries as autonomous decision-makers with real operational consequences. Yet no formal architecture exists that simultaneously addresses what kinds of agents exist in enterprise settings, how they transition through operational states, who may authorize them to act, and how their actions are traced for accountability. Existing frameworks address these questions piecemeal: cognitive architectures model reasoning without governance; orchestration frameworks implement coordination without formal authority models; digital-twin formalisms cover lifecycle without delegation semantics. This paper presents a formal architecture for governed enterprise AI agents organized around four interlocking contributions. (1) An identity-anchored taxonomy classifies enterprise agents as Independent, Delegated, Asset-Bound, and Embedded agents, each with a defined authority source, identity anchor, lifecycle coupling, audit evidence type, and characteristic governance failure mode. Task labels such as analytical agent, coding agent, customer-service agent, or compliance-review agent are treated as deployment scenarios rather than primitive agent forms. (2) A three-layer governance architecture structures each agent across a cognitive layer (L_C), a platform layer (L_P), and an organizational layer (L_O), with explicit inter-layer accountability flows. (3) A lifecycle finite-state machine (FSM) with seven states — Design, Approval, Deployment, Operation, Adaptation, Suspension, and Retirement — formalizes the governance obligations and permissible transitions at each phase, including mandatory human-in-the-loop checkpoints at Approval and Suspension. (4) A delegation relation δ maps human principals to agent authority scopes using form-conditioned source authority and a policy cap, so that multiple partial grants cannot aggregate into unauthorized privilege escalation and delegation chains remain bounded. The architecture is grounded by an accountability matrix that cross-references agent form against governance instrument, audit evidence, escalation path, and compliance artifact. Four enterprise deployment scenarios — independent customer/IT service, delegated data governance, asset-bound operations, and embedded software engineering assistance — instantiate the architecture and demonstrate its practical discriminating power across the four identity-anchored forms and multiple task modes. Series: ANANKE (AICT Series on Governed Agent Dynamics), Paper 2. Companion preprint (ANANKE-1): https://doi.org/10.5281/zenodo.20576283 == HTML VERSION (for Zenodo rich-text editor if needed) ==

Enterprise AI agents are increasingly deployed across workflows, platforms, and organizational boundaries as autonomous decision-makers with real operational consequences. Yet no formal architecture exists that simultaneously addresses what kinds of agents exist in enterprise settings, how they transition through operational states, who may authorize them to act, and how their actions are traced for accountability. Existing frameworks address these questions piecemeal: cognitive architectures model reasoning without governance; orchestration frameworks implement coordination without formal authority models; digital-twin formalisms cover lifecycle without delegation semantics.

This paper presents a formal architecture for governed enterprise AI agents organized around four interlocking contributions.

(1) Identity-anchored taxonomy. Enterprise agents are classified as Independent, Delegated, Asset-Bound, and Embedded agents, each with a defined authority source, identity anchor, lifecycle coupling, audit evidence type, and characteristic governance failure mode. Task labels such as analytical agent, coding agent, customer-service agent, or compliance-review agent are treated as deployment scenarios rather than primitive agent forms.

(2) Three-layer governance architecture. Every agent is structured across a cognitive layer (L_C), a platform layer (L_P), and an organizational layer (L_O), with explicit inter-layer accountability flows.

(3) Lifecycle finite-state machine. A seven-state FSM — Design, Approval, Deployment, Operation, Adaptation, Suspension, and Retirement — formalizes the governance obligations and permissible transitions at each phase, including mandatory human-in-the-loop checkpoints at Approval and Suspension.

(4) Delegation relation and accountability matrix. A delegation relation δ maps human principals to agent authority scopes using form-conditioned source authority and a policy cap, preventing unauthorized privilege escalation through grant aggregation. An accountability matrix cross-references agent form against governance instrument, audit evidence, escalation path, and compliance artifact.

Four enterprise deployment scenarios — independent customer/IT service, delegated data governance, asset-bound operations, and embedded software engineering assistance — instantiate the architecture and demonstrate its practical discriminating power across the four identity-anchored forms and multiple task modes.

Series: ANANKE (AICT Series on Governed Agent Dynamics), Paper 2. Companion preprint (ANANKE-1): https://doi.org/10.5281/zenodo.20576283