What concrete definition separates AI agents from chatbots?

Caleb Sponheim, writing for Nielsen Norman Group, argues that "agent" has become "a diffuse, ambiguous term that means everything and nothing" because marketing, sales, engineering, product and leadership each use a different working definition. NN/g's fix is a single concrete definition: "An AI agent is a system that pursues a goal by iteratively taking actions, evaluating progress, and deciding its own next steps." The goal can be narrow ("fix this failing test") or open-ended ("research competitors ... and summarize the findings"); what makes something an agent is that the goal shapes which actions it takes and how it judges progress, and that it builds on the result of each prior step. A plain chatbot fails this test: it "received your input, generated text, and stopped. There was no iteration, no action beyond producing a reply, and no self-directed next step."

The reasoning hinges on treating widely cited agent properties — perception, adaptation, memory — as consequences of the three-part loop rather than as separate checklist items. "Evaluating progress" already covers perception, since "a coding agent that reads its own error logs is perceiving its environment; so is a self-driving car processing lidar data." Adaptation "is implicit in deciding its own next steps," since a system able to choose a next step can in principle choose a different one after failure. Memory "is implicit in iteratively," since building on prior work requires retaining some record of it. Sponheim's stated reason for keeping these implicit is economy of definition: "a definition that requires perception, adaptation, memory, self-direction, the capability to act, and iteration risks becoming a checklist that generates more disagreement than clarity." The piece also stresses that no single piece suffices alone — an LLM "that reflects on its own output is performing a form of evaluation, but evaluation alone is not enough" — and illustrates the line with a real pair of responses to the same Algolia analytics task: a chatbot that described the steps and handed them back, versus an agent that navigated the dashboard, failed to change a date range, adjusted, and tried again.

This is a binary, loop-based test (a system either closes the act-evaluate-decide loop or it doesn't), which sits in tension with graded models of agency in the library. Should we define AI agents as a spectrum rather than a category? scores systems along four separate axes instead of drawing a line; NN/g's definition supplies exactly the line OpenAI declines to draw, though both treat current chatbots as falling short of full agency. Does machine agency exist on a spectrum rather than binary? likewise frames agency as graded (passive hammer to cooperative smart home); Sponheim's loop condenses that spectrum into a single qualifying threshold while still accommodating degrees of reliability above it. When should a workflow actually use an AI agent? presumes a working notion of "agent" to decide when one is warranted — this NN/g piece is the definitional groundwork a designer would need before applying that guide.

The excerpt is a design-practice definition aimed at giving UX professionals a shared vocabulary, not a technical specification or a claim about how reliably current systems adapt or retain memory in practice; Sponheim explicitly separates can decide a different next step from does so reliably, calling the latter "a question of usefulness, not of identity." It also does not address trust, safety or accountability once a system clears the definitional bar. The implication NN/g draws is narrower than a general theory of agency: the definition is meant to let practitioners classify a given system quickly, not to resolve how much perception, adaptation or memory makes an agent good.