Define the room before the assistant
A custom agent always answers from a context, whether that context is designed or accidental. A useful design begins by naming the room: the team, workflow and information the assistant is meant to operate within.
This prevents the blank-chat-box problem. Instead of asking what an agent could possibly do, ask what a person in this room needs, which sources they trust and what should remain outside it.
Decision one: which sources count
List the approved source systems and the reason each belongs. A policy library, project repository and public website may all be useful, but they carry different authority and update patterns.
Where answers should be inspectable, preserve a route back to the exact source. A citation is most helpful when a reviewer can open the underlying passage rather than merely see a document title.
- Source owner and update responsibility
- Freshness requirements and change monitoring
- Citation or evidence expected in the response
Decision two: who can bring what into the room
Shared and personal context need different rules. A group assistant should not silently make one person’s private material available to everyone else, and a team assistant should not cross client or project boundaries.
Write those boundaries as permission scenarios. Test what happens when a user asks for material outside their context, not only when they make an expected request.
Decision three: how the agent says it does not know
Unsupported requests are part of normal operation. Decide whether the assistant should say the source is missing, ask a clarifying question, direct the user to another workflow or involve a person.
A good refusal preserves momentum. It explains the boundary and offers a useful next action without presenting a guess as an answer.
Decision four: where human handoff goes
Handoff needs a destination and enough context for the receiving person to act. That may be an existing support queue, an operations owner or a project reviewer.
Carry across the request, relevant sources and the reason for escalation. The person receiving the handoff should not have to reconstruct the conversation before helping.
Evaluate scenarios, not a demo script
A polished happy-path demonstration cannot test these decisions. Build a small scenario set covering missing sources, conflicting material, out-of-scope requests, permission boundaries and handoff.
Review the whole response behaviour: answer, evidence, boundary and next action. Those scenarios become the basis for a system a team can operate with confidence.