Stop Using Agents for Deterministic Workflows
Not every workflow becomes better when an agent is allowed to choose the next step.
If the process is always validate input, check account status, calculate a fee, and update a record, ordinary code provides stronger guarantees at lower cost. A model introduces latency, variability, token costs, and new failure modes while adding little value.
Agents become interesting when the path depends on interpretation or exploration: gathering evidence from inconsistent documents, comparing possible explanations, or determining which questions remain unanswered. Even there, the agent should operate inside a bounded set of permissions and outcomes.
A hybrid workflow is often the better architecture. Use deterministic routing for known steps, a model for genuinely ambiguous decisions, and explicit validation at the transition back into system state.
Input -> Validation -> Fixed business rules
|
ambiguous evidence?
/ \
no yes
| |
commit Agent analysis
|
validate -> review
The engineering test is straightforward: list the decisions that require judgment. If you cannot name them, you probably do not need an agent. If you can name them, specify what evidence the agent may inspect and which actions remain forbidden.
Good automation removes uncertainty where it can. It does not manufacture uncertainty for the sake of using AI.
Top comments (0)