"Agent" became one of those words that gets attached to almost any AI feature. A product that summarises a document becomes an "AI agent." A chatbot with a search tool becomes an "agent platform."
The label matters less than the design choice underneath. How much does the model decide on its own? Does it follow a fixed sequence of steps you've designed, or does it choose what to do next?
In my work on Frontline at WiZR, most of our enterprise workflows ended up as fairly structured sequences with the model doing specific jobs inside them. That wasn't for lack of ambition. Enterprise customers wanted predictable, reviewable output, and structure gave them that.
The difference
A workflow is a set of steps you design. The model might draft text in step two and classify it in step four, but the order and the logic are fixed in code.
An agent decides which steps to take, often in a loop. It chooses tools, reads the results, and decides what to do next until it thinks the task is done.
Anthropic's post on building effective agents draws this same distinction and recommends starting with the simplest approach that works, adding autonomy only when it's needed. 1
How they compare
| Factor | Workflow | Agent |
|---|---|---|
| Predictability | High. Same steps every time | Lower. Path can vary |
| Debugging | Easier. You know which step failed | Harder. You trace what it decided |
| Cost | Easier to estimate | Can vary a lot per task |
| Latency | More predictable | Can grow with more loops |
| Flexibility | Limited to designed paths | Can handle unexpected tasks |
| Review | Easy to put checkpoints between steps | Harder to know where to check |
When a workflow is the better choice
- The task follows a fairly consistent pattern.
- Mistakes are costly, so you want checkpoints.
- Cost and response time need to be predictable.
- Users or customers need to understand how output was produced.
Most of the automation candidates I'd pick first fit this description. Drafting standard documents, summarising, reformatting and classifying are all well suited to workflows.
When an agent might make sense
- The task is genuinely open-ended and the steps can't be predicted.
- There are many possible tools and the right combination depends on the situation.
- Occasional mistakes are acceptable, or a person reviews the final result.
- The value of handling unusual cases is high enough to justify the extra cost and unpredictability.
Research tasks, some troubleshooting tasks and complex multi-step requests are examples where agents can do things a fixed workflow can't.
The middle ground
Most real products end up somewhere in between. A mostly fixed workflow, with one step where the model has some freedom to choose a tool or decide whether more information is needed. That gets you some flexibility without giving up control of the whole process.
Questions I'd ask before building an agent
- Could a workflow with a few branches handle most cases?
- What happens when the agent picks the wrong tool or loops too long?
- How will we limit cost per task?
- How will we test it, given that its path varies?
- Where does a person review the result?
If those questions don't have good answers yet, I'd start with a workflow and learn from real usage. You can always give the model more freedom later. Taking it away after users depend on it is harder.