Agent or fixed workflow: what is the difference?

A workflow follows a route that someone wrote in advance. An agent picks its own route as it goes. Anthropic's wording: a workflow runs models and tools through predefined code paths. In an agent, the model steers its own process and tool use¹.

An automation with an AI step in the middle is therefore usually a workflow. The AI does one job, such as sorting a message or drafting a reply, and the fixed steps around it decide what happens next.

The test: who chooses the next step

Ask one question of any product. At each point, what decides what happens next? That's the test.

• If a script or a rule decides, it's a workflow, however clever the AI step.
• If the model decides which tool to use and when to stop, it fits the agent description.
• If both are involved, it's a mix, and the makers treat that as normal.

Anthropic calls the whole family "agentic systems" for this reason¹. Its blog post notes that some customers use "agent" for fully autonomous systems, and others use it for more prescriptive set procedures¹.

Named workflow patterns

Anthropic names several workflow patterns, each with fixed structure¹:

¹

The last one is close to the agent end. The central model decides the subtasks as it goes, so they aren't fixed in advance. Anthropic lists other patterns too, such as running steps in parallel, and one where a call generates a response while another gives evaluation and feedback in a loop¹.

Why the difference matters to a buyer or user

Anthropic says workflows offer predictability and consistency for well-defined tasks¹. It also says agentic systems often trade delay and cost for better task performance¹. Its advice to developers is to find the simplest solution possible and add complexity only when needed. That might mean not building an agentic system at all¹.

So a fixed workflow isn't a lesser product. For a job with the same steps every time, predictability is the point. What matters is that the label matches the thing. A tool sold as an agent that is really three fixed steps has different risks from one that picks its own actions. The risks of the second kind are on the page about mistakes and loops.

A third meaning

OpenAI's documentation uses "agent" for something that can plan and complete tasks with tools, work with other agents and keep context². It also offers a toolkit where your own application controls the agent loop². So a business can build an agent whose loop its own software supervises.

Wikipedia's wider definition stresses acting autonomously toward goals³. All three sources point at the same thing. How much the system decides for itself is the question. Whether the product wears the word "agent" is not.