Not every multi-step flow is an agent
A workflow can contain many steps and still be deterministic. If the sequence, inputs, and permitted actions are known in advance, regular application logic is usually easier to test, cheaper to run, and simpler to explain.
An agent becomes useful when the system must choose among tools, adapt its plan, interpret unstructured context, or decide when the goal has been met. Those capabilities create value, but they also increase the number of ways a run can fail.
A practical selection test
Before introducing an agent, ask whether the workflow has all four of these properties.
- Meaningful ambiguity that cannot be expressed as stable rules.
- A bounded set of tools with clear inputs, outputs, and permissions.
- A measurable definition of success or an escalation condition.
- Enough value per completed task to justify variable latency and cost.
Use the least autonomy that works
A useful progression is deterministic function, model-assisted function, single agent, and only then coordinated agents. Each step adds flexibility and a larger test surface. Starting with the smallest workable design makes failures easier to localize.
This is not an argument against agents. It is an argument for giving model judgment a precise job instead of making it the default control plane for the whole application.
Approval gates protect the business process
Actions that communicate externally, change money, update records of consequence, or disclose sensitive data deserve explicit controls. Human approval is not evidence that the system failed to automate. It is a deliberate boundary around the decisions a business is not ready to delegate.