A new arXiv survey tries to map every corner of agentic AI reasoning, and the result reads like a catalogue of how unsettled the field still is.
Researchers organized agentic reasoning into three layers: foundational reasoning, covering core single-agent skills like planning, tool use, and search in stable environments; self-evolving reasoning, where agents refine those skills through feedback, memory, and adaptation; and collective multi-agent reasoning, which covers coordination and shared goals across multiple agents. The paper also separates in-context reasoning, which scales performance through prompting and orchestration at run time, from post-training reasoning, which shapes behavior via reinforcement learning and fine-tuning. It reviews frameworks used in science, robotics, healthcare, autonomous research, and math. It closes with a list of open problems, including personalization, long-horizon interaction, world modeling, scalable multi-agent training, and governance for real-world deployment.
This is a taxonomy, not a product launch, but the framing is useful precisely because "agent" has become marketing shorthand for anything with a tool-calling loop bolted on. Separating single-agent tool use from genuinely coordinated, collective multi-agent systems gives buyers and builders a sharper vocabulary for what a given "agentic" product actually does.
The most telling part of the paper is its own to-do list: the authors name scalable multi-agent training and real-world governance as unresolved, not solved. Worth remembering next time a demo shows a swarm of agents working in perfect harmony.