Software given the loop, not just the answer, perceive, decide, act, repeat, without waiting for a human at every step.
The formal definition of “agent” in AI research predates the current chatbot era by decades. Michael Wooldridge and Nicholas Jennings’ foundational 1995 survey defines an agent by autonomy, reactivity, and pro-activeness: a system that senses its environment and acts on it to satisfy its own objectives, without being driven by a human decision at each step.
What’s new in the current wave is not the concept but the substrate: large predictive models (see Predictive Technology) are now fluent enough to be wrapped in an agent loop — read a goal, decide a next action, call a tool, observe the result, decide again — and marketed as “AI agents.” The underlying reasoning is still prediction, chained; the agency is architectural, not a new kind of cognition.
This is the mechanism behind orchestrating agents versus being orchestrated by them, who designs and supervises that loop, and who is simply subject to its output. See Agency for the vault’s broader framework on that divide.
Source: Wooldridge, M. & Jennings, N.R. (1995), “Intelligent Agents: Theory and Practice,” The Knowledge Engineering Review, 10(2), 115–152.