Choosing a workflow over an agent
“When would you choose a workflow over an autonomous agent?”
What this tests
- Bias toward determinism when it is sufficient
- Concrete criteria rather than intuition
- Understanding of operational consequences: testing, cost, latency, debugging
- Ability to design hybrids
Answers by level
Read the beginner answer first and notice what is missing.
I choose a workflow whenever I can enumerate the steps and their order ahead of time, which is most business processes: extract fields, validate, look up, decide, act, notify. Each step is a node with a typed input and output; some nodes are LLM calls, most are code. See Workflow State Graph and Business Process Automation.
The reasons are operational. A workflow has bounded cost and latency, each node is unit-testable, intermediate state is inspectable, retries are local to the failing node, and a wrong output can be traced to a specific step. An agent offers flexibility I do not need if the steps are known, and charges for it in variance and debugging time.
When one step genuinely needs open-ended reasoning, I keep the workflow and make that node a bounded agent with its own iteration cap and tool set. The rest of the system stays deterministic.
Green flags · Red flags
- Criterion: can the steps be enumerated ahead of time
- Lists concrete operational benefits: bounded cost, unit-testable nodes, local retries
- Distinguishes content uncertainty from sequence uncertainty
- Proposes a hybrid with a bounded agent node
- Backs the choice with eval results and cost/latency numbers
- Chooses agents for "complex" tasks without defining complex
- Believes agents handle edge cases better by default
- No mention of testing or cost
- Cannot describe a hybrid