Agentic AI Systems: A Taxonomy of Autonomy, Delegation and Accountability for Tool-Using Language-Model Agents

Abstract

The term agentic AI is now applied to systems that range from a chatbot permitted to call a calculator to a persistent, self-directing orchestrator that spawns sub-agents, browses the open web and executes irreversible actions in production environments. Treating these as a single category obscures the fact that each step along that range introduces a distinct failure mode and a distinct accountability question. This paper proposes a taxonomy that makes the steps explicit. Drawing on the recent survey literature on language-model agents, on empirical work on tool-use and function-calling reliability, on the security literature on prompt injection in tool-integrated systems, and on the philosophical and legal literature on delegation, responsibility gaps and meaningful human control, it defines six levels of autonomy for tool-using language-model agents (responder, tool-calling assistant, bounded executor, goal-directed agent, orchestrating agent and persistent autonomous agent) and three cross-cutting dimensions along which any agent at any level can be positioned: the action surface it can reach, the environment it is exposed to, and the oversight regime under which it runs. For each level the paper identifies the characteristic failure modes and the accountability instruments that remain feasible. Its central argument is that accountability does not diminish as autonomy rises, as the delegation framing implies, but must be deliberately scaled up with it, and that most current deployments invert this relationship by expanding autonomy faster than they expand the instruments by which a human can answer for what the agent does. The paper closes with seven design principles for bounded autonomy, a note on deployment under infrastructure constraints, and a research agenda.