What a workflow is and what an agent is
A workflow is a chain of steps that a person set: a request came in, the model identified the topic, the request went to the right manager. A model (a neural network that understands text) may sit inside a workflow, but the order of steps does not depend on it.
An agent chooses its next step and tool itself. This is exactly how Anthropic divides agentic systems: in workflows the steps are set by code, and in agents the model decides what to do next.
A workflow is predictable: on the same input it does the same thing, it is easy to check and cheap to run. In Anthropic’s words, workflows give predictability and consistency where the task is clearly defined.
Builders such as n8n can do both: in one workflow some steps are fixed, and at the right point an AI agent decides by itself. So you choose not a program but a form of solution for each step.
An agent is more flexible, but each of its steps is a new request to the model and a new chance to make a mistake. How an agent goes through its cycle step by step is covered on the page about how an AI agent works.
How an agent differs from a chatbot with buttons, we wrote in our blog. Here the comparison is with another neighbor: a workflow that already has a model but makes no decisions.
Five signs of a task for an agent
The signs come from two guides by model developers: Anthropic (December 2024) and OpenAI (April 2025). The more matches, the more sense an agent makes.
1. The number of steps is unknown in advance. Anthropic recommends agents for open-ended tasks where you cannot predict how many steps will be needed and cannot hard-code the path.
2. The decision depends on context and exceptions. OpenAI’s example: whether to approve a refund to a customer in support.
3. The rules have grown and are costly to maintain. OpenAI gives the example of vendor security reviews, where each rule update costs money and creates errors.
4. Lots of free text: emails, documents, chats with a customer. OpenAI’s example: processing an insurance claim from a customer’s statement.
5. The result can be checked. In support it is a resolved request, in programming it is passed automated tests: this is how Anthropic explains why agents took hold exactly there.
If one sign matched or none, start with a workflow. OpenAI writes the same: if a task does not meet these criteria, a solution with rules set in advance may be enough.
Check the signs on your own task in writing: list the last ten cases and mark where the steps matched and where they were different each time. If the steps are almost always the same, a workflow will handle it more cheaply.
The cost of a wrong choice
An agent where a workflow would do costs more and makes more mistakes. Anthropic warns that an agent’s autonomy means higher costs and a risk of compounding errors, so agents are tested in an isolated environment and given guardrails.
The opposite mistake also costs money. A workflow for a task with exceptions grows branches, and each new rule touches the old ones; OpenAI calls such systems unwieldy, and their updates costly and error-prone.
There is a third cost too: time. A workflow is built and checked faster, because its behavior is visible from the diagram. An agent has to be tested on dozens of real cases, since its path may be different each time.
Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027 because of costs, unclear value and weak risk control. A wrongly chosen form of solution hits the first two reasons at once.
So the choice between an agent and a workflow is made before the build, by the task and its economics. Why this matters more than the choice of tool is covered in a separate article.
Examples from sales and support
Support, agent. Anthropic calls support a natural task for an agent: a conversation with a customer needs outside data and actions. The agent finds the customer’s data, order history and knowledge base articles, processes a refund or updates a ticket, and success is measured by a resolved request.
The same source notes that some companies charge for such agents only for successfully resolved requests. This is an example of paying for the result, not for hours of work.
Support, workflow. An email with the subject invoice goes to accounting, one with the subject delivery goes to logistics, and the customer gets a confirmation. The model here only identifies the topic; the rules do the rest.
Sales, workflow. A request from the website goes into the CRM, the customer gets an email with the price list, and the manager gets a task to call. The steps are always the same, so no agent is needed here.
Sales, agent. A customer writes in free text that they need two hundred boxes by Friday, with delivery. The agent sorts out the items, checks stock and prices, asks a clarifying question and creates a deal; the number of steps depends on the customer, which matches signs 1 and 4 from the list above.
A combination often works well: a workflow accepts the request and sorts it, and an agent steps in only for complex cases.