Choose tools by their effects

Start with two capabilities: look up an order and save a draft. Treat these as example interfaces to implement in your chosen application.

Validate tool inputs outside the model. Reject missing identifiers and requests outside the permitted account. A failed lookup is a tool error, not evidence that the order does not exist.

Keep credentials in the integration that calls the service. The agent only needs the result it is allowed to see.

Separate working context from memory

Working context is what the agent needs for this task: the message, lookup result, and draft. Longer-lived memory is information deliberately retained for future tasks.

For this worker, keep approved tone guidelines and a link to the current support policy. Fetch delivery status each time; yesterday’s status is not a reliable answer today.

For each retained fact, record its source, when it was checked, who can correct it, and when to delete or refresh it. Do not turn every customer conversation into permanent memory. Never store access tokens in conversational notes.

Hand back a decision, with context

Use a consistent handoff when the order is missing, sources disagree, or the customer requests an exception.

Needs review: delivery status is unavailable.
Checked: order 1042; lookup returned a timeout.
Prepared: a draft acknowledging the delay without inventing an ETA.
Next decision: retry the lookup or assign the ticket to support.

A handoff should preserve work without claiming success. Repeated failures should end the run and reach the owner rather than create a retry loop.