The actual distinction
- Workflow: you define the steps; AI does the parts requiring comprehension
- Agent: the model chooses which tools to use and in what order, until it decides it is done
That difference matters enormously for reliability, cost and how you debug a failure.
Why workflows usually win in business
| Property | Workflow | Agent |
|---|---|---|
| Predictability | High | Variable |
| Cost per run | Known | Variable, sometimes surprising |
| Debugging | Straightforward | Considerably harder |
| Auditability | Clear | Requires effort |
| Handling novel situations | Poor | Better |
Businesses value the first four more than the fifth, almost always. That is why the fashionable answer and the right answer differ here.
Where agents genuinely help
- Research tasks where the next step depends on what was just found
- Diagnosis, where the questions to ask depend on the answers so far
- Long-tail cases too varied to enumerate as steps
Note that all three are exploratory rather than transactional. That is the pattern.
Constrain agents heavily
- A small, explicit set of tools rather than broad capability
- A step limit and a cost limit per run
- No write access to anything consequential
- A full trace of every step, retained
- A human checkpoint before anything irreversible
The hybrid pattern
A workflow for the known path, with an agent handling the specific step that is genuinely open-ended, inside strict limits.
That gets the predictability where it matters and the flexibility where it helps, which is what most real systems end up looking like.