A useful AI agent starts with its boundaries

Start with the decision boundary

The first question is not what the agent should look like. It is which decisions the system may make, which information it may use and what output is expected.

Separating information, action and recommendation makes evaluation clearer and prevents uncertain material from being presented as fact.

• Which requests can the agent handle alone?
• Which sources can it use?
• Which actions require approval?
• Which cases must be handed over immediately?

Treat sources as part of the product

A knowledge assistant depends on a defined collection, access rights, versions and someone responsible for updates.

When evidence is absent or contradictory, abstaining is an expected behaviour—not an error to hide.

Design human handoff as a real journey

A generic “contact support” message is not a handoff. The destination, transferred context and next expected state should be clear.

The person taking over should receive the useful history without forcing the user to repeat the exchange.

Evaluate cases, not impressions

A smooth demonstration is not sufficient evidence. Representative, ambiguous and refusal cases form a reusable evaluation set.

That set can be run again whenever sources, rules or model behaviour change.