Agentic Workflow
Most useful automation sits somewhere between a rigid script and a free-running agent. An agentic workflow is that middle ground: a multi-step process where a language model does the thinking at each step, while the shape of the process — which steps exist and how they connect — is designed in advance by a human.
The shape of it
A workflow is a graph the developer draws; an agent is a loop the model drives. In an agentic workflow the model's judgement is used at specific, bounded points and the control flow around those points is ordinary code. The recurring shapes:
- Prompt chaining — each step's output feeds the next, with a programmatic check in between.
- Routing — a classification step sends the task down one of several prepared branches.
- Parallelisation — independent subtasks run at once and a final step merges them.
- Orchestrator–worker — a planning step splits the task and delegates the pieces.
- Evaluator–optimiser — a generating step and a critiquing step loop until a quality bar is met.
Why you would choose one
A defined graph is predictable, cheap to test, and bounded by construction: you know the maximum number of model calls before you run it. A fully autonomous loop is more flexible and much harder to reason about — it might take twelve steps or ninety, and its failure modes are open-ended. The honest rule is to use the least autonomy the task needs, and to promote a step to open-ended reasoning only when its branches cannot be enumerated. LangGraph exists largely to make that graph explicit, and most production "agents" turn out to be workflows with one genuinely agentic node.
An agentic workflow is a task pipeline in which model-driven agents make the routing and tool decisions at each stage, rather than following a sequence fixed in advance. Steps branch, retry, or loop based on intermediate results. The term marks the middle ground between rigid automation and full autonomy: the pipeline constrains the overall path while the agent decides how to complete each step within it.
Think of a recipe versus a chef. A recipe lists the steps in order: chop, fry, simmer, serve. A chef handed "make dinner" decides everything themselves.
An agentic workflow is a recipe where a smart helper handles each step. A person lays out the stages in advance — read the email, work out what it is about, draft a reply, check it — and the AI does the thinking inside each stage. It is not free to invent a different plan.
That trade is deliberate. You give up some flexibility and gain something valuable in return: you know roughly what it will do, how long it will take, and what it will cost.