Mental BoundA workflow is an AI system where the sequence of steps is decided in advance: the code calls the model at fixed points, runs tools, and passes results from one stage to the next. The model supplies the intelligence inside each step but never chooses the route. In Anthropic's framing, this predefined path is what separates a workflow from an agent, which directs its own process and selects its own tools.
Common patterns include prompt chaining (a fixed sequence of model calls, each building on the last), routing (classify a request, send it down the right path), parallelization (independent pieces run at once), orchestrator–workers (one model splits the work, others execute), and evaluator–optimizer (one drafts, another critiques).
Workflows suit well-understood, repeatable tasks: they are predictable, testable, and cheaper to run. Agents are reserved for problems where the path cannot be known in advance.