Separate the interface from the execution pattern
Anthropic distinguishes predefined workflow paths from agents that choose their next steps and tool use. That distinction describes control of the process; it does not tell you whether a particular product can publish, send or modify data.
Match the approach to a concrete task
For a one-off rewrite of approved product facts, a reviewed conversation may be sufficient. A recurring report with stable input fields may fit a predefined workflow. An investigation that must choose different evidence depending on what it finds may benefit from agent-led steps. Compare these approaches on the same tasks, including difficult and incomplete inputs, rather than choosing by the most impressive demonstration.
Request evidence of the complete task
For a collection review, inspect the exact source URL and date, query scope, product facts and proposed changes. Ask what happens when a source is unavailable or contradicts another. If the demo claims publication, inspect the actual destination and revision; a generated preview is only a draft. Confirm permitted actions and a stopping condition before any live trial.
Keep the acceptance test independent of the label
Prepare several representative tasks with expected facts, disallowed changes and a named reviewer. Record factual errors, unsupported conclusions, manual corrections, elapsed time and operational cost. Repeat after meaningful system changes. A successful example does not establish consistent performance; a workflow that needs substantial checking may still be useful if its total benefit is demonstrated.
Three ways to prepare a collection update
Illustration: a conversation rewrites supplied copy; a workflow fetches agreed reports and produces a fixed brief; an agent investigates a missing collection signal using permitted tools. All three still need the correct collection, verified facts and an accepted output. None earns publishing permission simply because it is called an agent.
Use this at work
- Specify the task and available tools.
- Observe how the next step is selected.
- Test missing inputs and conflicting facts.
- Verify the result at its actual destination.
Reference material
Sources checked 4 October 2026. Definitions are attributed where cited; the retail review methods and examples are Faccelerate editorial proposals, not product capability claims.