A system can use a powerful language model without needing the model to control the whole process. The right amount of autonomy depends on how much the next step genuinely changes from one task to another.

Anthropic’s architectural distinction between workflows and agents is straightforward: workflows follow predefined paths, while agents dynamically direct their processes and tool use. That is a useful starting point for evaluating products that share the same “agent” label.

When the path is already known

Take an illustrative document-routing process. A file arrives, required fields are checked, the request is classified and the result goes to the appropriate queue. A model may help interpret an unfamiliar description, but the surrounding steps can remain explicit.

Giving the model freedom to invent a different route may add little value when the organization already knows where each request belongs. The important work is defining categories, handling ambiguity and making exceptions visible.

When the next step depends on discovery

A different task might ask an assistant to investigate why a customer cannot complete onboarding. The relevant evidence could be spread across a support history, an integration status and an account configuration.

Here, the sequence may depend on what the assistant finds. A missing credential suggests a different next check from a rejected permission request. Limited agent-led investigation may therefore be useful, provided the system knows when to stop and what it may change.

These examples illustrate architectural choices. They are not results from a comparative product test.

Separate investigation from authority

A team can allow flexibility in gathering evidence while keeping corrective actions tightly controlled. An agent might inspect approved sources and suggest a fix, then hand the decision to a person before modifying production settings.

This matters because the ability to choose a helpful next step is different from the authority to carry it out. Our analysis of website access barriers explores a related limit: understanding a task does not establish permission to complete it.

Ask vendors to show the boundary

  • Which steps are fixed, and which does the model choose?
  • What evidence causes the workflow to change direction?
  • What prevents repeated attempts from continuing indefinitely?
  • Which actions need human approval?
  • How can the team reproduce and investigate a failure?

Choose the simplest design that meets the need

Our recommendation is to start with a clear process and introduce flexibility where real cases demonstrate a need for it. A fixed workflow can still deliver meaningful automation. An agent can still work within strict boundaries.

The decision is not a contest between old and new technology. It is a question of where judgment improves the result enough to justify additional monitoring, evaluation and recovery work.