A new academic survey says the word agent in AI products is not just marketing fluff. It actually maps to a specific software architecture.
Researchers combed through vendor documentation and academic papers to catalog seven recurring forms of software built around large language models: chatbots, custom agents, retrieval-augmented generation (RAG, where a model pulls in outside documents before answering), AI-enhanced workflows, copilots, coding agents, and a hybrid called agentic RAG. They compared each form across four traits: its architectural pattern, who controls execution and when a human can step in, how many times the model gets called per task, and how it uses external tools. Across 22 real systems, the label copilot consistently meant a router-and-worker setup that pauses for step-by-step user approval, while the newer label agent marked software where the AI plans and runs several steps on its own, showing the user only the final result. Coding agents from four major providers, despite different branding, all turned out to run the same loop: reason, act, then delegate to subagents.
That matters because agent has become the single most overused word in AI marketing, slapped on everything from spreadsheet macros to fully autonomous systems, which makes it nearly impossible for buyers to know what they are actually paying for. This survey is evidence that at least some of that vocabulary is doing real engineering work rather than just dressing up a chatbot. Vendors converging on the same control pattern independently is a more convincing signal than any press release.
Still, a term being architecturally meaningful today is no guarantee it stays that way once a product manager needs a new word for next quarter's roadmap slide.