Agentic Workflow

Most software runs the exact same steps every time, in the exact same order. An agentic workflow works almost like that. The one difference: at a step or two, an AI model makes the decision instead of the code.

Think of it as sitting between two other things. On one side is a fully fixed workflow, where the model never decides anything. On the other is a full AI agent, where the model decides every step itself, in whatever order it wants. An agentic workflow is the middle ground: fixed steps, with a model-made decision at one or two of them. It's one specific point on the agentic AI scale, not a separate idea from it.

What makes it a workflow, not a full agent

The number of decision points and where they sit in the process are fixed ahead of time in code, even though what the model decides at each one isn't. A step that classifies an input always runs at the same place and always produces one classification — the code decides what happens after each possible answer. A full agent has no such fixed shape: the model can also decide how many decisions to make and in what order, not just what to decide at a predetermined point.

Two common patterns

This shape has names. Anthropic's engineering guidance describes several fixed-shape-with-a-decision-point patterns like this; two of them come up often enough to know by name:

• Routing: the model classifies an input, then code sends it down one of several predetermined paths, each with its own handling. The set of possible paths is fixed; which one gets used isn't.
• Prompt chaining: a task is broken into a fixed sequence of steps, where each step's output feeds the next. The model reasons at each step, but the number and order of steps is set in code, not decided by the model.

Both come from the same source already cited on the AI agent page.

Routing a support ticket

A support-ticket system reads each incoming ticket and has the model classify it as billing, technical, or a refund request. Each category has its own fixed next steps — billing tickets get a templated response and escalate automatically if the amount is over a threshold; technical tickets trigger a documentation search before a reply is drafted; refund tickets get checked against policy before approval. The three paths and what happens on each are written in code ahead of time. The model's job is narrow: decide which of the three categories a given ticket belongs to.

When this is the right choice

When you can list the possible paths or steps ahead of time, even if you can't predict every input that will come through them. An agent takes on risk across its whole sequence, one decision at a time, all the way through. This pattern only takes on that risk at the one or two points you actually chose — every other step runs the same way it always does, untouched by the model's judgment.

When you need a full agent instead

When the paths themselves can't be enumerated in advance, or the number of steps needed genuinely depends on what's discovered along the way — the debugging example on the AI agent page is exactly this case. A fixed set of routes can't cover a problem where you don't know ahead of time how many investigative steps it will take to find the cause.