Skip to content

Agentic AI, AI agents and workflows: when you need which

18 Nov 2026

·

7 minute read

A workflow is software in which code decides every step and a language model does some of the work inside those steps. An AI agent is software in which the model decides the next step itself, choosing which tool to call and when to stop. "Agentic AI" is used for both, and for larger systems of several agents with memory, depending on who is writing. Because the sources disagree, the label is a poor guide when you are buying or building. The useful question is narrower: at each step, who decides what happens next, the code or the model? Our own blog pipeline answers it differently for different steps, and the reasons are a fair guide to when you need which.

People are searching for the difference

We pulled Google Trends data on 14 November 2026 for a keyword study. Worldwide, searches for "ai agents" ran at about three-fifths the level of searches for "robotics", and grew over the year. In the UAE, "ai agents vs agentic ai" and "gen ai vs agentic ai" were breakout queries, Google's label for its fastest-growing searches, and "generative ai vs agentic ai" roughly quadrupled. In India, "agentic ai course" was among the top related queries for "agentic ai".

Those are people trying to sort out vocabulary before they spend money or time on it. The vocabulary does not settle.

The definitions disagree, so use a test instead

Anthropic's widely read guide, Building effective agents, uses "agentic systems" as the umbrella. Under it sit workflows, "where LLMs and tools are orchestrated through predefined code paths", and agents, "where LLMs dynamically direct their own processes and tool usage". On this reading a fixed pipeline that calls a model is agentic.

A 2025 review titled AI Agents vs. Agentic AI draws the line somewhere else entirely. It describes AI agents as modular, task-specific systems built on language models, and reserves agentic AI for systems marked by multi-agent collaboration, dynamic task decomposition, persistent memory and coordinated autonomy. On that reading, a single agent that answers support tickets is not agentic AI at all.

Both are reasonable. They cannot both govern a contract. So when someone describes a system as agentic, or as an agent, ask three things:

  1. At each step, does code or the model decide what happens next?
  2. Is the number of steps known before the run starts?
  3. Can the model take an action with consequences, such as sending, paying, deleting or publishing, without a person or a hard rule in between?

The answers tell you what you are dealing with, and what can go wrong, in a way the label never will.

Our blog pipeline is a workflow with bounded agent steps

The system that drafts posts for this site makes a good worked example, because we made the choice step by step and wrote down why.

StepWho decidesLimit
Should anything run today?Code: a pre-flight scriptStops on no approved topic, two drafts waiting, or seven posts already queued
Which topic?Code: first approved topic whose material existsThe model never picks
Which sources to read?The modelAt most 8 research calls and 5 sources, from an approved list
What to write?The modelOnly facts from its material files and its sources
Is the draft acceptable?Code: the quality gateFixed rules, no model
How to fix a failing draft?The modelAt most two attempts, then the topic is marked blocked
Publish?A person: merging the pull request reaches staging, and the owner's publish step reaches the live siteThe run cannot publish in its default mode

By Anthropic's definition the whole thing is an agentic system built as a workflow. By the 2025 review's it contains no agentic AI at all. By our three questions, code decides three of the seven steps, the model decides three and a person decides one, and every step the model decides runs inside limits set in code.

Where a fixed path beat an agent

Choosing the topic is the clearest case. A model asked "is there anything worth writing today?" will usually find something, and on most days the answer we want is nothing. The pre-flight script answers from the topic list and exits in about a second when nothing is approved, at a cost of a fraction of a cent. An agent would have spent a research session deciding the same thing, and some days it would have decided wrong.

Checking the draft is the second. A model reviewing its own draft for generated-sounding phrases finds some and misses others, and it can be persuaded that a borderline sentence is fine. A script with a list of 50 banned phrases and 11 banned sentence shapes cannot be persuaded. We describe that gate in detail in our post on the pipeline.

Publishing is the third, and it stays with a person. The one time a batch of posts passed every automated check and still fell short, it was the owner reading them who caught it. No step that makes a draft public is left to the model.

Where the model's judgement earns its place

Research and writing are where the steps cannot be listed in advance. Which source answers this question? Is the material file enough, or does the post need an official document to anchor a claim? How should a section be rewritten when the gate rejects a sentence for a pattern? These are open-ended in exactly the way Anthropic describes as suiting agents: you cannot predict the number of steps or hard-code the path.

So those steps run with the model in charge, with limits around them. The research budget caps the cost. The approved-source list caps what it can cite, enforced by the gate. The fix-up loop caps the retries. The model makes real decisions, and the expensive outcomes stay out of its reach.

When you need which

You haveBuildBecause
A task whose steps you can list in advanceA workflowIt is cheaper, faster, and fails in ways you can reproduce
A task with a known shape but open-ended partsA workflow with bounded agent stepsAutonomy where it helps, code where mistakes are costly
A task where the path depends on what each step findsAn agent, with budgets, approvals and logsHard-coding the path is impossible
Several such tasks that must coordinateThink twice before multiple agentsCoordination failures are the main way these systems break

The last row deserves its own warning. A system of several agents with shared memory is what the 2025 review calls agentic AI, and it is where most failure reports come from. We went through the published failure modes in why multi-agent systems fail. Anthropic's advice points the same way: find the simplest solution possible, and add complexity only when it is needed, which "might mean not building agentic systems at all".

Questions to ask before you buy something called agentic

  • Which steps does the model decide, and which are fixed in code? Ask for the list.
  • What is the most consequential action it can take without a person approving it?
  • What stops a run that loops, and what is the spending limit per run?
  • Where is every tool call logged, and who reads the log?
  • If the model were replaced tomorrow, which steps would behave differently?

A vendor who can answer those clearly has built something you can operate. If the answers are vague, the label on the box was doing the work. If you are deciding between a workflow and an agent for a real process, we are happy to look at it with you.

AI tools

Ask ChatGPTAsk Claude

Start with the problem that matters