Concept guide

What is agentic AI? A plain-English guide with examples

What agentic AI means, how an AI agent loop works (plan, act with tools, observe, repeat), where agents fail, and what to learn to build one.

2 min readFree, no sign-up

Short answer: agentic AI is software in which a language model does not only answer a question but pursues a goal over several steps: it decides what to do next, calls tools (search, code, databases, other APIs), looks at the result and continues until the goal is met or it should stop. The word "agent" refers to that loop, not to a special kind of model.

The loop at the centre

Almost every agent is some version of this cycle:

  1. Goal: the user states what they want ("find the three cheapest flights and put them in a table").
  2. Plan or choose: the model decides the next action, such as calling a search tool.
  3. Act: the system runs the tool with the arguments the model produced.
  4. Observe: the tool result is added to the model's context.
  5. Repeat or stop: the model either takes another step or gives the final answer.

You can read the terms in our glossary: AI agent, agent loop and agentic workflow.

A small worked example

A support assistant receives: "Where is my order 4821?" A plain chatbot can only guess. An agent can call lookup_order(4821), read that the parcel is "out for delivery", then call get_policy("delivery-delay") if the customer asks about a refund, and finally answer using both results. The model wrote the plan and the tool arguments; your code ran the tools and enforced the rules.

Workflow or agent?

Many useful systems are workflows: fixed steps written by an engineer, with the model filling one or two of them. A true agent lets the model choose the steps. Workflows are more predictable and cheaper; agents handle open-ended tasks but need stronger limits. A good rule: use the simplest design that solves the task, and add autonomy only where the task cannot be written as steps in advance.

Where agents go wrong

  • Wrong tool or wrong arguments: the model picks a tool it should not, or invents a parameter.
  • Loops that never end: without a step limit and a stopping rule, an agent can repeat itself and run up cost.
  • Lost context: long runs overflow the context window, so memory and summarising matter.
  • Unsafe actions: an agent that can send emails or spend money needs approval gates and narrow permissions.

How to judge an agent

Test it on realistic tasks and score the outcome, not only the final text: did it call the right tools, in a sensible order, within the step and cost budget? See agent evaluation and our guide to building an LLM evaluation pipeline.

What to learn to build one

Python, calling a model API with structured outputs, tool (function) calling, retrieval (RAG), logging and traces, evaluation, and guardrails (LLM guardrails). For several agents that cooperate, read multi-agent systems.

Words used on this page

AI agent · Agent loop · Agentic workflow · Agent memory · Agent evaluation