Agentic AI describes AI systems that pursue goals through multiple steps, often using planning, memory, tools, environmental observations, and stopping rules. Some systems coordinate several specialized agents.
An agent is not only a language model. Reliability depends on the closed workflow: permissions, tool behavior, evidence, evaluators, memory, human review, recovery, and monitoring. Errors can be corrected or amplified through feedback.
Example
A research agent searches sources, stores findings, asks another component to draft, and invokes a publishing tool after evaluation. If unverified summaries enter shared memory, apparent agreement can reinforce error.