Design a Phone Agent That Books Repair Visits for a Property Management Company

Design a phone agent for a property management company. The company's customers call the agent to say which services they need, for example to book a time for someone to come and make a repair, and the agent handles these requests during the call. The interviewer treated this as designing the agent itself, not as a large-scale system design exercise. Only this prompt was reported, so settle the scope first. ```hint Start from the calls List the main kinds of call the agent must handle and, for each, the information it has to collect before it can act and the system it has to act on. ``` ```hint Know what the agent must not decide alone Decide which situations need a person, and how a call reaches that person without the caller having to repeat everything. ``` ### Constraints and Clarifications - Not given: the list of services, the systems the agent can read and change (tenant records, work orders, technician schedules), languages, call volume, office hours, or how emergencies are handled. Ask, or state your assumptions. - Focus on the agent's behavior, tools, safeguards and evaluation rather than on scaling infrastructure. ### Clarifying Questions - Who calls: tenants only, or also property owners and prospective renters? How is a caller identified and verified? - Which requests are in scope: repair and maintenance visits only, or also rent, lease or amenity questions? - Which systems hold work orders and technician schedules, and may the agent create and change bookings directly? - How should urgent problems, such as a gas smell, flooding or no heat, be handled, and what happens outside office hours? - What does success mean for the company: calls resolved without a person, visits booked correctly, caller satisfaction? ### What a Strong Answer Covers - Scope: the call types, the callers, the systems the agent acts on, and what success means - A conversation design with the state the agent must fill (who, where, what problem, how urgent, access, time window) and a confirmation before acting - The agent architecture: speech recognition, a language model with tools for lookup and booking, speech synthesis, and how latency and interruptions are handled - Guardrails: emergency triage and escalation, no invented availability or promises, identity checks before sharing or changing data - Failure handling for tool errors, misheard details and callers who go off script - Evaluation before launch and monitoring after, with concrete metrics ### Follow-up Questions - A caller reports water coming through the ceiling late at night. Walk through exactly what the agent does. - The scheduling system is down during a call. What does the agent say and do? - How do you test a new prompt or model version before it handles real calls? - The same tenant calls three times about the same leak. How does the agent recognize this, and what changes?