What is the difference between an AI agent and an AI workflow?

In this article:

• What was the client's challenge?
• What is an AI agent and what is an AI workflow?
• When do you choose a workflow over an AI agent?
• What does an AI workflow actually deliver?
• How is AI changing the engineer's role?

What was the client's challenge?

The client is a large education group with multiple programs and brands. Course materials are the core of what they offer, and those materials are still largely created by hand. A single piece can involve four or five roles, with a lot of reading and research, and a turnaround time of weeks.

The client had already experimented with AI, but with mixed quality. A familiar pattern: many AI pilots end up delivering no demonstrable value. So the question they brought to Blis was concrete.

Can AI contribute to better content, in less time and at lower cost?

Blis built an AI solution that takes over tasks within those roles and shortens the process to days. The roles themselves didn't disappear. Robert: "The client deliberately wanted to keep the control moments, the human in the loop. They don't trust AI enough for that yet."

Agent, workflow, orchestrator. These words are often used interchangeably, even though the difference determines what you build and how much control you keep along the way. So first the concepts themselves, before we get to the choices made in this project.

What is an AI agent and what is an AI workflow?

The difference comes down to who determines the route. With an AI agent, you give it a goal and let go of how it gets there: the agent chooses its own path. With a workflow, you determine the route: which steps, in what order, and what happens when something is uncertain.

Think of an agent as a digital colleague. You give it a role with clear instructions and quality standards, and it carries out that task independently. An agent like this consists of three things: a role (who it is and what it needs to do), tools (what it's allowed to use) and a brain (the language model that thinks for it).

Robert: "The simplest form of a tool is web search. You basically tell it: go find out what's relevant content for this course material."

And just like a new colleague, an agent needs to be onboarded. You do that with context: examples of good and bad work, and guardrails like tone of voice and quality guidelines.

An AI workflow chains several of those agents together, each with its own task. The output of one becomes the input of the next, and you set the order and the checkpoints. An agent is a specialist that chooses its own path. A workflow is the process where you link those specialists together.

There's a third form, and it makes the difference sharper: the orchestrator agent. That's a single agent that's allowed to run through the entire flow itself and decides along the way which steps are needed.

Robert: "With an orchestrator, you give a specific agent the freedom to run through that flow itself."

A workflow and an orchestrator aim for the same result. The difference is freedom. With a workflow, the path is fixed. With an orchestrator, the agent chooses the path. And that exact choice became important in this project.

When do you choose a workflow over an AI agent?

If a process is tightly defined and has fixed control moments, you choose a workflow. When we started the project with the education organization, we initially thought we needed an orchestrator agent. It turned out to be a workflow after all.

Robert: "The process is put together so tightly that we shouldn't deviate from it. The validation steps in particular were so important that a workflow was the obvious choice."

The problem with an orchestrator is exactly its strength. It gets to decide for itself. An agent convinced of its own judgment might decide a review step can wait this time. With course material that multiple people review and approve at different points, that's not an option. The control moments, the human-in-the-loop, weren't a side detail. They were the entire point.

What does an AI workflow actually deliver?

In this project, a manual process that took weeks was reduced to days, with a 24-step workflow producing well over 120 pages of course material. That result didn't come from the technology alone.

We never do this with engineers alone. In an earlier episode we called it the golden triangle: the AI engineer, the business consultant and the client's product owner building together. Each brings something the others don't have.

Robert built the technology and deliberately let go of the content.

Robert: "I set things up, because I don't know enough about the client's work process. I just want to make sure it works technically and that the output is generated correctly."

Blis's business consultant handled the other half. They talked with the client to surface the real process: which steps are involved in creating course material, and where a human needs to review it. Based on that, the consultant determined how many agents were needed and what role each one had. The prompts Robert had drafted were completed by the consultant with domain expertise Robert didn't have himself.

That conversation also revealed how big the process really was. Robert started with a six-step workflow that produced four pages. Once the real process came to light, it grew to 24 steps, resulting in those 120 pages.

The work that used to be the core, the coding, is no longer where the time goes. That time now goes into knowing what you're building and setting up the process correctly. And that knowledge comes exactly from the triangle: the AI engineer builds, the business consultant and the product owner determine what's good.

How is AI changing the engineer's role?

The engineer writes less code themselves and directs more. This was Robert's first project that he built entirely with Claude Code, without making any changes in the editor himself.

Robert: "I'm much more the orchestrator of the entire development process."

At Blis, we look at the lead engineer's role the same way. They work at the system level, on the architecture and how everything fits together, while AI handles the implementation. The work shifts to what we call the middle loop: the layer between writing code yourself and delivering the result. That's where you steer the AI and catch the results that sound plausible but aren't correct.

You see it in the small choices. When Robert switched from open source tooling to Microsoft Foundry, he ran both versions side by side for a while. Same end result, 20,000 fewer lines of code. He even outsources testing. He gives Claude the URL and the passwords and has it test the complete 24-step workflow end to end in the browser, while he watches with one eye.

The models keep getting better, but they still make choices that make Robert think: that's not how I want it. That's where he steers.

Robert: "I set the boundaries the model needs to stick to, and then we start building."

The developer doesn't disappear. Their role shifts. From writing code to guarding the bigger picture and stepping in at the right moments.

👉 Listen to or watch the full episode of Blis Bytes