In plain terms
Think of an assembly line where each station is staffed by a model. One station classifies the request, the next drafts a reply, a third checks it. The line itself never changes; only the work at each station is done by AI.
Why it matters
Most production AI that works today is a workflow, not a free-roaming agent. Fixed paths are cheaper, faster, easier to test and easier to explain to an auditor. Start here and add autonomy only where the fixed path demonstrably fails.
Example
Incoming support email: step one labels it (billing, technical, complaint), step two drafts an answer from the knowledge base, step three checks tone and policy. If the check fails, the draft goes back once; otherwise it lands in a support agent's queue.
Most often confused with
Agentic Workflow vs. Agent
In a workflow the path is fixed: you can draw it as a flowchart before it runs. An agent chooses its path at run time, so two runs of the same task may take different routes. Workflows trade flexibility for predictability.
Under the hood
Common patterns: prompt chaining (each call consumes the previous output), routing (classify, then send to a specialised branch), parallelisation (fan out and aggregate), orchestrator–workers, and evaluator–optimizer loops. Because control flow lives in code, each step can be evaluated, cached and retried separately. Many systems are hybrids: a deterministic workflow with one agentic step inside a tightly scoped sandbox.