agent workflow

An agent workflow refers to the whole process by which AI finishes a task not in one shot, but by moving through a series of ordered steps. Think of cooking: prepping ingredients, cooking, then plating, each step in sequence. For an AI that helps with coding, the stages might break down into planning, writing code, checking the screen, and reviewing the results.

Instead of one chef doing everything from start to finish, different specialists can be called in at each stage. One AI might handle planning, another might write the code, and yet another might do the review. By assigning AIs with different strengths to different stages, you can build a combination that's better on cost or performance than relying on a single all-purpose AI.

Recently, developer tools have started supporting this kind of setup, which means teams aren't locked into one AI and can adjust the combination depending on the situation.

In articles, you'll see phrasing like "different open models can be assigned to each stage of the agent workflow." Here, each stage refers to specific tasks like coding, planning, vision, and review. One common misunderstanding is that agent workflow isn't the name of a specific product — it refers to the structure, or method, by which AI processes a task.

Try asking an AI chatbot to break a request into stages like this, and you'll get a feel for how a workflow actually operates:

"Split the following task into three stages — plan, execute, review. In stage 1, just come up with a plan for how to do it. In stage 2, actually carry it out. In stage 3, check your own results. Task: [write your task here]"

Comparing the output from a one-shot answer versus a staged request makes the concept of a workflow click quickly.