Agent, or workflow?
I keep seeing agents where a workflow would do. Teams hand a model steps that should be plain, deterministic automation, and make predictable work nondeterministic in the process. Reasoning earns its place in some of the work, not all of it. Answer seven questions and I will tell you which parts of this actually need an agent, and which should stay a workflow you can test.
The executable version, a Claude skill that reads a real spec, is on GitHub. This is the design gate in the two gates. If the work lives on Salesforce, switch the lens to Agentforce and the same verdict renders in Flow, prompt templates, topics, and actions.
What does it mainly do?
How much of the work follows fixed steps or rules?
How many distinct jobs does it span?
How reversible are its actions?
Where does its knowledge live?
What are the latency and scale needs?
How much should it act on its own?
Answer the questions and I will tell you which parts of this actually need an agent, and which should stay a workflow you can test.