agents medium complexity mvp
Workflow
A workflow is a controlled sequence of steps where the application decides the path.
Decision
Use a workflow when the product needs reliability, predictable state, and a known process.
Use when
- Form processing
- Approval flows
- Support routing
- Repeatable business operations
Avoid when
- Open-ended exploration
- Tasks requiring adaptive planning
- Research with unknown intermediate steps
- Cases where the next action cannot be predefined
Why workflows win early
Most AI products should start with workflows. They are easier to test, easier to explain, and easier to recover when something fails.
An LLM can still be useful inside a workflow: classify, extract, summarize, draft, or choose from constrained options. The difference is that the application owns the process.
When workflows are too rigid
A workflow struggles when the task path is unknown. If each observation changes what should happen next, an agent may be a better fit.
Common mistakes
- Using an agent where a three-step workflow would be safer.
- Letting the model control state transitions without validation.
- Forgetting fallback states and manual review.
Next decision
Use workflows by default for MVPs. Move toward agents only when the workflow becomes too brittle for the task.