AI Agents vs. Agentic Workflows: Which One Do You Need?

“Agentic” gets attached to almost everything that calls a language model more than once. That is unhelpful, because two quite different designs hide behind the word, and picking the wrong one is the most common reason an AI project is slower, costlier and less reliable than it needs to be.

Two designs, one distinction

A useful distinction, popularised in Anthropic’s guide Building effective agents, is this:

In a workflow, LLMs and tools are orchestrated through predefined code paths. You decide the sequence of steps in advance; the model fills in each step.
In an agent, the LLM dynamically directs its own process and tool use, deciding at each step what to do next.

In a workflow the control flow lives in your code. In an agent it lives in the model’s decisions, which buys flexibility and costs predictability.

Neither is better. Workflows trade flexibility for predictability, speed and lower cost. Agents trade the other way. Most production systems are mostly workflow with a small amount of agency in the places where it pays off.

Five workflow patterns worth knowing

1. Prompt chaining

Split a task into a fixed sequence, where each model call works on the output of the last, optionally with a programmatic check between steps. Example: draft a product description, check it against a style guide, then translate it. Use it when the task decomposes cleanly and you want each step to be easier than the whole.

2. Routing

Classify the input first, then send it down a specialised path. Example: a support message is classified as billing, technical or general, and each category has its own prompt and tools. Use it when different inputs genuinely need different handling.

3. Parallelization

Run several model calls at once and combine the results. Two flavours: sectioning, where independent sub-tasks run in parallel, and voting, where the same question is asked several times to raise confidence. Use it for speed, or when you want several independent opinions.

4. Orchestrator–workers

A central model breaks the task into sub-tasks that you could not list in advance, hands them to worker models, and merges the results. Example: a change that touches an unknown set of files. It looks agentic because the plan is dynamic, but the structure is still yours. We go deeper in multi-agent systems.

5. Evaluator–optimizer

One model produces a result, another critiques it, and the loop repeats until the critique is satisfied. Example: literary translation, or code that must pass a linter. It works when there are clear criteria and iteration measurably improves the answer.

When you genuinely need an agent

Reach for a true agent when you cannot predict the steps in advance, the number of steps is open-ended, and you have some way to check progress. Typical cases: fixing a bug in an unfamiliar codebase, researching a question across many sources, or running a scientific calculation where the next experiment depends on the last result. (See agentic AI for science.)

A decision guide

Question	If yes	If no
Can you write down the steps in advance?	Workflow	Consider an agent
Is a wrong or surprising result costly?	Workflow, or an agent with approvals	An autonomous agent is more acceptable
Is latency or cost tightly limited?	Workflow	Agent is affordable
Can you verify progress automatically (tests, validators)?	Agent becomes practical	Keep a human in the loop
Does the input vary wildly in kind?	Routing, then specialised workflows	One workflow is enough

Start simple, add agency last

The most reliable advice is also the least glamorous. Begin with a single model call and a good prompt. Add retrieval or a tool only when you can show it helps. Add a fixed chain when one call is not enough. Introduce an agent loop only when the problem needs open-ended decisions, and when you do, give it a step budget, good logs and a way to stop. Each layer you add makes the system harder to debug, so every layer should earn its place.

New to the topic? Start with What is Agentic AI?