A new survey paper takes the sprawling, overhyped world of AI agents and tries to sort it into something coherent.
The paper proposes a unified formal language for classifying LLM-based agentic reasoning frameworks, then uses it to split the field into three tiers of increasing complexity: single-agent methods, tool-based methods that let an agent call outside resources, and multi-agent methods where several agents coordinate on a task. It then surveys how these frameworks actually get used, across scientific discovery, healthcare, software engineering, social and economic modeling, and general-purpose tasks. Along the way, it compares how each category gets evaluated, since judging an agent that calls one tool is a different problem than judging a group of agents negotiating toward an answer.
Right now, "AI agent" gets stretched over everything from a chatbot with web search to systems where multiple model instances debate each other before producing output. A shared three-tier vocabulary gives researchers and engineers a way to actually compare those systems instead of arguing past each other about what counts as agentic.
It is a map of the territory, not a verdict on which framework wins - so it won't tell you what to build, just what to call it once you have.