Beyond the hype.

Workflow or AI agent: who decides what happens next?

A workflow follows a sequence planned in the code. An agent can choose the next step itself. That difference changes a lot when it comes to cost, testing and control.

In this article

Your application queries an external service to complete a customer record. That service goes down. First scenario: the program tries a second source, as planned by the developer. Second scenario: the model looks for another way to get the information by itself. That is a very concrete difference between a workflow and an agent.

The criterion: who decides the next action

What separates the two is not the presence of AI, nor whether the system seems intelligent. It is who decides the next action. In a workflow, the path is set by the code. In an agent, the model can choose part of the path.

The distinction is not ours. Anthropic clearly separates workflows, which orchestrate models and tools through predefined code paths, from agents, which dynamically direct their own process. Simon Willison sums up an agent as a system that runs tools in a loop to achieve a goal.

In practice, three questions are enough: who decides the next action, the number of attempts, and when to stop? If the answer is “the code”, you have a workflow. If it is “the model”, an agent. For the difference between chatbot, copilot and agent, see AI agent, behind the word, what are we talking about.

What is not enough to make an agent

  • Self-evaluation. An AI that scores its own output stays within a fixed frame, as long as it does not choose what comes next.
  • An evolving knowledge base. Updating rules or data does not give the model control of the sequence.
  • An “adaptive” system. Adjusting to an input changes the result, not necessarily the path.

A system can also combine the two: a workflow overall, with a single step left to the model's initiative. That is often the healthiest setting.

What autonomy costs

When the model is in charge, outputs become less predictable, the path taken is harder to trace and explain, and both validation and security need more attention. Anthropic actually recommends adding autonomy only when it visibly improves the result.

Extra autonomy brings flexibility, but it also makes testing and control harder. If your steps are known in advance, a workflow may well be enough.

Author · M.Baudoux - Software development & AI consulting

Sources

GDN Journal · Beyond the hype.All articles

Get the next GDN Journal articles

News explained, in-depth analysis and real-world experience in AI, cybersecurity and software engineering.

One email per new article. Nothing else.

By subscribing, you agree to receive GDN Journal by email whenever a new article is published. Your address is processed by GDN and its email provider, Mailjet. It is never shared. Privacy policy.