Design Customer Support with Virtual-Agent-to-Human Escalation

Design a customer-support system in which a user converses with a virtual agent and can be escalated to a human support agent. Explain how conversation context, ownership, and message delivery are preserved during the handoff. ### Constraints & Assumptions - The source supplies the virtual-agent-to-human support scenario, without scale, channel, staffing, or SLA details. - **Practice scope:** a text conversation has durable messages, one current response owner, an escalation queue, and a human-agent interface. Escalation may be requested by the user or triggered by an explicit routing policy. - Clarify any bot tools and permissions separately; the handoff must not silently repeat an external action already taken by the virtual agent. - Human availability is limited. The system needs a visible state for waiting rather than pretending every escalation is immediately staffed. ### Clarifying Questions to Ask - What triggers escalation, and can the user always request it directly? - Which skills, language, priority, or availability rules choose a human agent? - Can a user continue sending messages while waiting, and what may the bot do during that period? - Which context may be shared with the human agent, and how are generated summaries distinguished from the original transcript? ### Part 1 — Conversation and Ownership Model Describe message persistence, virtual-agent responses, escalation requests, queueing, human claim, and resolution. #### What This Part Should Cover - A durable conversation state machine with one current response owner. - An explicit waiting state and idempotent escalation request. - Message identity and ordering across reconnects and repeated sends. ### Part 2 — Transfer Context and Stop Conflicting Replies Explain the handoff boundary, transcript and summary delivery, messages arriving during transfer, and a delayed bot response after human takeover. #### What This Part Should Cover - A versioned ownership transition and fencing of stale bot or agent writes. - Original messages and tool outcomes available alongside any generated summary. - No lost messages between the transferred snapshot and live updates. ### Part 3 — Recover and Operate Describe human disconnects, reassignment, cancellation, and operational measurements. #### What This Part Should Cover - Durable queue and assignment state with controlled recovery. - No silent conversation abandonment or duplicate ownership. - Metrics for waiting, handoff correctness, and resolution without invented performance claims. ```hint Handoff is a state transition, not just a notification Sending a transcript to a human does not stop an in-flight bot response or prevent another human from claiming the same conversation. ``` ### What a Strong Answer Covers - A complete virtual-to-human lifecycle with durable history and clear ownership. - Consistent context transfer and fencing of stale responses. - Visible waiting and recovery behavior when no human is immediately available. ### Follow-up Questions - What should happen if the bot completes a tool action after escalation was requested? - How can a human inspect what the bot actually did rather than relying only on its summary? - How would a reassignment avoid losing messages sent while the previous human was disconnected?