Agentic AI refers to AI systems that can autonomously pursue multi-step goals by planning sequences of actions, invoking external tools, and adapting based on intermediate results—all without requiring a human to direct each individual step. Unlike a single-turn question-answering model, an agentic system maintains a task objective across many interactions, decides which tools to call (web search, code execution, file management, APIs), and revises its approach when earlier steps fail or produce unexpected outputs.

Think of it like the difference between asking a colleague a single question and delegating an entire project to them. You hand over the goal; they figure out the steps, gather resources, encounter obstacles, and report back when done—or ask for clarification only when genuinely stuck.

A simple agentic loop might look like this:

while not task_complete:
    action = llm.decide_next_action(goal, history)
    result = tool_executor.run(action)
    history.append((action, result))
    task_complete = llm.check_completion(goal, history)

Real frameworks such as LangChain, AutoGen, and Anthropic's Claude tool-use API implement more robust versions of this loop, including error handling, memory management, and parallel sub-agent spawning.

Agentic AI draws heavily on Foundation Model capabilities, Prompt Engineering to structure the agent's reasoning, Chain of Thought for intermediate planning, RAG for retrieving external knowledge, and RLHF to shape safe decision-making. AI Alignment concerns are especially acute here because autonomous action amplifies any misalignment in the model's goals.

You will encounter agentic AI in product release notes for models with tool-use capabilities, in framework documentation for orchestration libraries, and in AI Safety literature discussing how to maintain human oversight over systems that can take real-world actions.