WHERE TO START DESIGNING AN AGENT?

Don’t start by writing the System Prompt.

Start by describing the job the Agent will take on.

1. What job does the Agent exist to solve?

If you can’t describe the task in one or two clear sentences, the Agent’s scope may be too broad.

2. What is the real input?

Files, questions, data, forms, an API, or a state from another system?

3. What result counts as done?

A report, a decision, structured data, an alert, or an action?

4. How does a skilled person do that job?

Identify the method, the checklist, and the steps that must not be skipped.

5. What does the Agent need to know?

Identify what must become Context.

6. What does the Agent need access to?

Identify the Tools and data sources that are truly necessary.

7. What is the Agent not allowed to decide on its own?

Identify the boundary, human review, and high-risk actions.

8. How will the result be checked?

Identify verification, test cases, and acceptance criteria.

Don’t design an Agent from a prompt.
Design an Agent from the job, the method, and how the result is verified.