Agentic AI

How it works

A working agent has four components. A planner breaks a goal into ordered steps and decides what comes next. A tool layer executes those steps against real systems, holding whatever permissions that requires. A memory records what has already been attempted so the agent does not loop. An evaluation harness tests behaviour before a change reaches production.

Only the planner talks to the language model directly. That is the single most useful thing to understand about the category, because it explains why swapping models rarely fixes an agent that does not work.

A worked example

An invoice-matching agent receives an invoice, looks up the purchase order, compares line items, flags a discrepancy, and either posts the match or routes it to a human. Five steps, four of them tool calls. If the purchase order lookup fails, the agent has to decide whether to retry, wait, or escalate. That decision is the agentic part. A chatbot asked the same question would describe what should happen.

Often confused with

Why the distinction matters commercially

An agent holds permissions in your systems and takes actions with external effects. That makes it a security and operations question, not only an engineering one. A wrapped prompt sold as an agent will pass a demo and fail its first production incident, because nobody built the rollback path or the evaluation suite. Naming which one you are buying in the contract is the cheapest protection available.