Skip to content
Back to glossary

Agentic workflow definition

An agentic workflow is a multi-step process where a language model handles some of the steps, such as drafting, classifying, calling tools or checking output. The order of steps is fixed in code. That separates it from an AI agent, which plans its own steps.

The term gets used loosely, and in product marketing it can mean almost anything with a language model inside. The clearest definition comes from Anthropic's engineering guide Building effective agents (December 2024), which splits agentic systems in two. Workflows are systems where models and tools are orchestrated through predefined code paths. Agents are systems where the model dynamically directs its own process and tool use.

In a workflow, a developer decided the steps in advance and the model handles the parts that need language understanding. An agent decides the steps as it goes, which makes it more flexible and harder to predict.

What are the common agentic workflow patterns?

The Anthropic guide names five:

  • Prompt chaining: each model call works on the output of the previous one, with programmatic checks in between.
  • Routing: a first call classifies the input and sends it to a specialised next step.
  • Parallelization: several calls run at once, either on separate parts of the task or on the same task so the answers can be compared.
  • Orchestrator-workers: one model breaks the task down and hands the pieces to worker calls.
  • Evaluator-optimizer: one call produces a result, another critiques it, and the loop repeats until it passes.

Andrew Ng popularised the phrase "agentic workflow" in 2024 with a different cut of four patterns: reflection, tool use, planning and multi-agent collaboration. The two lists overlap more than they disagree.

Agentic workflow vs AI agent: which one do you need?

Usually the workflow. If you can write the steps down, write them down in code. A fixed path costs less to run, and when it fails you can find the failing step in a log. Anthropic's own advice is to start with the simplest thing that works, often a single well-prompted model call, and add steps only when they measurably improve the result. An agent is worth its cost when nobody can know the steps ahead of time, like tracking down a bug in an unfamiliar codebase.

What does an agentic workflow look like on a website?

Alt text for CMS images is a good fit, because the task is narrow and a person can check every result. A workflow for it could run like this:

  1. An editor uploads an image, and the upload event triggers a serverless function.
  2. The function sends the image to a vision-capable model and asks for alt text in each site language.
  3. Plain code checks the length and rejects answers that start with "Image of".
  4. The text is saved as a suggestion, and the editor accepts or rewrites it before publishing.

Every step is decided in advance, and the model only does the part that needs eyes and language. Nothing reaches the live site without a person. If the model starts writing bad alt text, you know which step to look at.