Agents vs. workflows: when you don't need an agent
A decision tree. Most "agentic" features are a state machine with better marketing.
A workflow is a path you defined: these steps, in this order, with a model call at some of them. An agent decides its own path — which tool to call, how many times, when to stop.
Agents buy flexibility and cost you determinism. Latency becomes unpredictable, spend becomes unbounded, and evaluation gets much harder because there is no single correct trace to compare against.
The honest test: can you draw the steps on a whiteboard? If yes, build the workflow. Reach for an agent only when the number of valid paths is genuinely too large to enumerate, and even then, bound it with a step limit and a budget.
The shape of it
- 01
Known steps
Workflow. Cheaper, faster, testable.
- 02
Known steps, unknown order
Workflow with a routing model call at the top.
- 03
Unknown steps, bounded tools
Agent with a hard step cap and a spend cap.
- 04
Unknown everything
Not a product yet. Narrow the problem.
Want this applied to your situation?
Sessions are direct and specific — you leave with a decision, not a reading list.
Book a session