Almost every AI conversation we have had this year starts in the same place. Someone has seen a demo of an agent that plans, calls tools, and works through a task on its own, and they want that inside their product. It is a reasonable thing to want. The demos are genuinely impressive.
Then we get into the detail of what the product actually needs to do, and most of the time the answer is not an agent. It is three or four specific AI steps wired into an ordinary workflow, with normal code doing everything in between. That version is less exciting to describe and considerably more likely to still be switched on in six months.
The distinction that matters is who decides the order of operations
An agent decides its own sequence. You give it a goal and some tools, and it works out which tool to call, in what order, and when it is finished. A workflow with AI in it does the opposite: your code owns the sequence, and the model is called at specific points to do a specific job, like classifying an email, extracting fields from a document, or drafting a reply.
That single difference drives everything else. In a workflow you can test each step in isolation, log exactly what happened, and fix a bad output by improving one prompt. In an agent, a wrong answer can come from any of a dozen decisions the model made along the way, and reproducing it is often the hardest part of the fix.
Where the agent shape genuinely earns its keep
We are not against agents. We build them. They make sense when the path really cannot be known in advance and the task is exploratory: searching across systems where the next query depends on what the last one returned, investigating a support issue where the useful next step varies case by case, or research work where a human would also be improvising.
Two conditions have to hold. The task needs a wide range of legitimate paths, so a fixed sequence would genuinely miss things. And the cost of a wrong step needs to be low or reversible, because the agent will take wrong steps. If a mistake sends a message, moves money, or writes to a customer record, the sequence should be yours, not the model’s.
What people are usually asking for when they say agent
In practice, when a founder says they want an agent, they usually want one of three things. They want the product to handle a whole task end to end instead of assisting with part of it. They want it to work with unstructured input rather than a rigid form. Or they want it to feel conversational rather than like filling in fields.
All three are achievable in a workflow. End to end just means covering every step, not that the model chooses the steps. Unstructured input is an extraction problem. Conversational is an interface decision. Building it as a workflow gets the same product outcome with failure modes you can actually debug at 6pm on a Friday.
Cost and latency decide more of this than teams expect
An agent loop makes an unpredictable number of model calls. A workflow makes a known number. On a feature running thousands of times a day, that difference shows up on the bill and in the response time, and it shows up unevenly, because the hardest cases loop the most. We have seen agent implementations where the worst five percent of requests cost twenty times the median and took a minute to return.
If the feature has to feel instant, or the unit economics need to be predictable per customer, that alone usually settles the design.
Our default, and when we move off it
We start with the smallest thing that produces the outcome, which is nearly always a workflow with a few well-scoped AI steps. We ship it, watch where it fails against real data, and then widen the model’s autonomy only at the points where the fixed sequence is demonstrably the limitation. Moving from workflow to agent later is straightforward. Moving the other way, after an agent has shipped and nobody can explain its behaviour, is a rebuild.
The honest summary is that agents are a real capability that is improving quickly, and also that most products asking for one need something narrower and more reliable first.
If you are working out what shape your AI feature should be, see how we work.