What is agentic AI? Workflows, agents and when to use each

Agentic AI is software in which a language model does more than answer a prompt. It calls tools, takes several steps and acts on what each step returns: it searches, reads, runs code, updates a record. It comes in 2 kinds. In a workflow, your code sets the path and the model does the steps. In an agent, the model decides its own next step from what it has found so far, until the job is done. Use the simplest kind that does the job. Most tasks need a workflow, and only open-ended ones need an agent.

That split comes from Anthropic's guide to building agents, and it is the most useful way to think about the term. "Agentic" is less a product category than a dial: how much of the path the model chooses.

Agentic AI vs generative AI

Generative AI produces something in reply to a prompt: an email, a function, a picture. Agentic AI puts that same kind of model inside a loop. The model reads a goal, chooses an action such as "search for X" or "run the tests", a tool carries it out, and the result goes back to the model for the next decision.

So every agentic system has a language modelLarge Language ModelA model trained on a huge amount of text to predict the next token. Further training turns it into an assistant that follows instructions. at its centre, but most uses of a language model are not agentic. A chatbot that answers from what it knows is generative. The same chatbot, once it can look up your order, issue a refund and email you the receipt, is agentic.

Workflows: the model does the steps, your code sets the path

Anthropic names 5 patterns. Each is an agentic workflowAgentic WorkflowA fixed sequence of model calls and tools that your code controls, such as classify, then look up, then draft. An agent decides its own next steps; a workflow follows the path you wrote.: predictable, testable and cheaper than an agent.

• Prompt chaining. Fixed steps in order, each model call working on the last one's output, with checks in code between them. Draft, check the length, translate.
• Routing. One call classifies the input and sends it down one of a few paths, each with its own prompt and tools. A refund question and a bug report should not share a prompt.
• Parallelization. Independent parts run at the same time and your code joins them, or the same task runs several times and the answers are compared.
• Orchestrator and workers. A model splits the task into parts it decides on, hands each to a worker call and combines the results. The parts vary, but the shape stays the same.
• Evaluator and optimizer. One call writes, another grades the result against criteria, and the first revises until it passes or runs out of tries.

Agents: the model chooses the steps

An AI agentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. runs a loop: look at the goal and what has happened, pick a tool, read the result, repeat. Nobody wrote the path down. That is what lets a coding agent fix a bug it has never seen, because it can open files, run tests and change course when a test fails.

It is also what makes agents slower, dearer and harder to trust. Anthropic says plainly that agentic systems trade latency and cost for better results on hard tasks. An agent can take 30 steps where a workflow takes 3, and every step sends the whole conversation back to the model. It can also take a wrong step on its own. Agents need limits, which stopping runaway agent bills and human approval before risky actions cover.

How to choose: the simplest pattern that works

Start with one model call. If that cannot do it, use a workflow. Use an agent only for the part whose steps depend on what it finds. Answer 4 questions about a job and see where it lands:

Try itWorkflow or agent? Find the simplest pattern for a job

The steps are

The same every timeOne of a few known pathsFound along the way

Splits into independent parts

NoYes

You can say what good looks like

NoYes

A wrong step is costly

NoYes

AgentAgent

Goal↓→Model: choose next step↓→Tool runs↓→Result back to model↓→Done

↺ Not done: choose the next step

Example: Find why last night's pipeline failed and open a fix.

The questions do most of the work:

1. Can you write the steps down? If they are always the same, chain them. If there are a few known cases, route between them. If they depend on what turns up, you need a model to decide them.
2. Do the parts depend on each other? If not, run them in parallel, or let an orchestrator hand them out.
3. Can you say what good looks like? Then a second call can grade the first, which beats hoping one call gets it right.
4. What does a wrong step cost? Anything that sends, pays, deletes or deploys gets a person's approval, whichever pattern you chose.

Many real systems mix them. A support system routes each ticket, runs fixed workflows for refunds and password resets, and hands only the odd cases to an agent with read-only tools.

What makes agentic AI work in practice

Anthropic lists 3 principles, and they hold for any framework:

• Keep it simple. Every extra step is cost, time and another place to fail.
• Show the plan. An agent that says what it is about to do and why is far easier to debug and to trust.
• Design the tools as carefully as the prompt. A model can only act through its tools. Clear names, clear descriptions, small inputs and useful errors matter more than a clever prompt. The tool useTool UseGiving a model a list of things it can ask to have done, such as search the web or read a file. The model asks; the program around it does the work and hands back the result. concept covers how a model calls them.

Then test it like software. How to evaluate an AI agent shows how to build tests that catch failures before users do. Our free AI engineering course teaches the parts in order, from a single model call to tools and agents.