What is Agentic AI?

How it works

"Agentic" is an adjective: it describes a spectrum of autonomy, not one product. At the low end, an LLM follows a fixed workflow where code decides the steps and the model fills in each one. At the high end, one or more AI agents decide their own plan, choose tools, delegate to other agents and keep going for minutes or hours.

The building blocks are the same throughout: a capable model, often a reasoning model; tools exposed by function calling or the Model Context Protocol; memory; and an agent harness that runs the loop, manages context and enforces limits. Anthropic groups both workflows and agents under "agentic systems", which is a useful way to think about it.

Why it matters

• It changes what AI is for. Generative AI produces content when asked; agentic AI takes on tasks and outcomes.
• It needs governance. More autonomy means more ways to go wrong: wrong actions, runaway costs, data exposure and prompt injection. Practical controls include least-privilege tools, spending and step caps, human approval for irreversible actions, and full logs.
• Hype is high. Many products labelled "agentic" are scripted workflows with an LLM step. That is often the right design, but ask what the system actually decides on its own.

Example

Generative AI: "Write a reply to this customer complaint." Agentic AI: "Resolve this complaint." The agentic system looks up the order, sees it was delivered late, checks the refund policy, issues a voucher within its approved limit, drafts the reply, and escalates to a person if the customer asks for more than the policy allows.