Agentic AI refers to systems that can perceive, plan, act, and adapt toward goals over multiple steps, rather than responding in a single isolated inference. In formal terms, an agent maintains an internal state $s_t$, observes the environment $o_t$, selects actions $a_t$, and updates its policy or plan based on feedback.

Agentic AI draws from classical AI planning, control theory, reinforcement learning, multi-agent systems, and modern large language model orchestration. The contemporary version is often built around foundation models that can reason in natural language and interface with external tools.

Agentic systems are increasingly used in software engineering, scientific discovery, operations, robotics, and research workflows. Their power comes from composing reasoning with action, but that same power introduces new failure modes such as compounding errors, unsafe tool use, and misaligned long-horizon behavior.