Start with a concrete result

“Help with customer support” is too broad. Try: “Read a delivery question, look up the order, and prepare a reply for review.”

The input is one customer message. The output is a draft with an order reference and the delivery status it used. You can check both against the source.

Use a fixed workflow when the steps are predictable. An agent becomes useful when the next step depends on what it finds.

Follow the decision loop

Imagine a customer asks, “Where is order 1042?”

1. Read: identify the order number and the requested information.
2. Act: call a permitted order-lookup tool.
3. Observe: inspect the returned status and its update time.
4. Finish or ask: prepare a source-backed draft, or ask for the missing detail.

This is an example design, not a connected demo. A tool error should produce an unresolved task, not an invented delivery date. Set a maximum number of attempts so the agent cannot retry forever.

Keep authority separate from ability

Looking up an order and issuing a refund are different permissions. Our example needs the first; it does not need the second.

Give it access to one approved data source and a place to save drafts. Sending messages, changing orders, and paying refunds stay with a person. Enforce those limits in the tools and application; a sentence in a prompt is not access control.

A useful completion record contains the draft, the source it used, and anything still uncertain.