AI agent or workflow: how to choose

What changes when a system becomes an agent?

A workflow prescribes the order of operations. An agent selects some of its next actions while working. Anthropic uses this distinction in its engineering guide. The distinction is about control, not the number of models or the presence of a chat window.

For a document intake process, extracting fields and checking a fixed list of requirements may be enough. An investigation with unknown starting evidence may require a system to choose another source or revise a plan. These are illustrative tasks, not customer case studies.

Describe the task before selecting the architecture

Write down the input, the observable output and the person who accepts it. Separate producing a draft from taking an external action. A system that recommends a change needs different permissions from one that applies it.

Ask what happens when a source is missing, a tool fails or two documents disagree. If the same recovery step is appropriate every time, an explicit workflow can remain easier to inspect. If the next step varies, evaluate whether an agent actually improves the result.

• Name one outcome that can be checked independently.
• List required tools and the smallest useful access scope.
• Describe an unresolved result without disguising it as success.
• Choose a baseline that completes the same task.

Compare accepted outcomes, not impressive conversations

Build a small, representative evaluation set from permitted material. Include straightforward requests, ambiguous requests and cases where an honest refusal is correct. Keep the expected decision separate from the system being evaluated.

For each attempt, record whether the result was accepted, how much human correction was needed, its elapsed execution time and its cost. Compare these against the existing process on the same inputs. A fluent answer is evidence of presentation quality; it is not evidence that the task was completed.

Keep action boundaries visible

A pilot should make its authority understandable to the operator. Show which source supports a conclusion and distinguish a proposed action from an executed one. Preserve enough evidence to investigate failures without collecting unnecessary sensitive content.

Useful questions for a supplier include: who owns the result, which errors stop the run, and how can a person resume or reverse a change? The answers should be demonstrated through the actual application path, including a failed attempt.

What should a pilot deliver?

A reviewable pilot produces an evaluation report, examples of accepted and rejected outputs, a description of access boundaries and a list of unresolved failure modes. Agree on the acceptance criteria before inspecting the final demonstration.

More autonomous behaviour is not automatically better. Retain a simpler process when it delivers the required outcome with less operational burden. Explore a richer agent architecture when measured results justify the additional moving parts.