Pre-Publication V2. A class of AI systems designed to pursue defined objectives autonomously or semi-autonomously by processing, modeling, and reasoning over available information – including their its contextual awareness – selecting and executing actions, and adapting behavior based on feedback with minimal or intermittent human intervention per action. A defining characteristic of Agentic AI is goal oriented actions versus prompt oriented actions. Unlike single-turn generative AI interactions (one input, one output), which require follow-up prompting, a Agentic systems may operate across extended sequences of decisions and actions, often utilizing external tools, APIs, code execution, web access, and other resources to complete multi-step tasks. Agentic AI are designed or intended to perform on a spectrum from fully human-supervised workflows to largely autonomous operation, and can operate within their local computing environment and across external systems, with the degree of human oversight representing a critical variable in safety, accountability, and liability frameworks. Key properties include goal persistence, environmental responsiveness, tool use, memory across steps, and the capacity to spawn or coordinate with other agents in multi-agent architectures.
Deliberation Summary:
(1) Discussed similarity of separate entry for an autonomous agent. (2) Removed the comparative framing against single-turn interactions, and (3) reworked the technical phrasing so it would be understood by a general reader. (4) Identified the contrast between goal-oriented and prompt-oriented action as the central point of the entry and simplified the surrounding language. (5) Acknowledged that such systems may operate within a local machine, taking over functions otherwise performed by a person, as well as reaching external systems, and (6) Declined to note that a system may depart from its instructions in pursuit of an assigned goal as consensus that was editorializing. (7) Noted that the sentence on human oversight should be retained as the only language in the list touching accountability, as it signals, without resolving, that autonomy does not remove human responsibility. (8) Minor consistency and formatting corrections were also noted.
Pre-Publication V1. A class of AI systems designed to pursue defined objectives autonomously or semi-autonomously by perceiving environmental state, reasoning over available information, selecting and executing actions, and adapting behavior based on feedback with minimal or intermittent human intervention per action. Unlike single-turn generative AI interactions, agentic systems operate across extended sequences of decisions and actions, often utilizing external tools, APIs, code execution, web access, and other resources to complete multi-step tasks. Agentic AI exists on a spectrum from fully human-supervised workflows to largely autonomous operation, with the degree of human oversight representing a critical variable in safety, accountability, and liability frameworks. Key properties include goal persistence, environmental responsiveness, tool use, memory across steps, and the capacity to spawn or coordinate with other agents in multi-agent architectures.