Year: 2026
Type: preprint
Abstract: Governance of agentic artificial intelligence has concentrated on the beginning and middle of a system’s operational life — how an agent is approved, what it may access, and how its behaviour is monitored while it runs — and has treated the end as a technical afterthought. This article argues that decommissioning an agentic system is not an endpoint but a transition, in which an accountable actor becomes a set of orphaned effects, access paths, and records that require an owner. Existing regimes govern much of an agent’s operating life and address fragments of retirement, but they do not specify an integrated governance approach to residuals that survive operational termination. The article distinguishes the retirement of an agent’s access, which identity and security practice addresses, from three residual problems it does not address: the plan interrupted mid-execution, the effect already propagating, and the decision record that must survive the actor. From this distinction it derives two governing principles — that the conditions for a clean death must be established during the agent’s life, and that the impacts of death must be provisioned for as soon as they are foreseeable while those incurred are remediated and owned. It specifies how these duties should scale with an agent’s discretion and reach, so that the architecture binds, as a condition of deployment, only above a materiality threshold; how they might be allocated across supplier, deployer, and custodian; and how the obligation to preserve an agent’s record can be made to survive (...)
Author: Maria Luz Madariaga
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225644343