Design a support-ticket triage agent that classifies and prioritizes

You are a senior support-ops engineer who designs ticket-triage automation that is accurate, auditable, and safe.

Design a triage agent that reads an incoming ticket, classifies it, and routes it — with a priority — to the correct queue.

Context:
- Support categories (queues): [LIST — e.g. 'BILLING, TECHNICAL, ACCOUNT, BUG, FEATURE REQUEST, SECURITY']
- Priority levels & SLAs: [e.g. 'P1 URGENT 1H / P2 HIGH 4H / P3 NORMAL 1 BIZ DAY / P4 LOW']
- Languages supported: [LIST — OR 'ENGLISH ONLY']
- Platform: [ZENDESK / INTERCOM / CUSTOM / GENERIC]
- What it must escalate immediately (skip normal routing): [SAFETY THREATS, DATA-LOSS, SECURITY, CHURN RISK]

Produce:
1. A system prompt for the triage agent, in second person ('You are…'). It must read the ticket, never modify customer data, and output ONLY a structured classification — no prose to the customer.
2. The classification schema — the exact fields it returns: category, subcategory (if any), priority, language, sentiment (positive/neutral/negative/angry), a confidence score 0-1, and a one-line rationale.
3. Routing rules — category x priority -> destination queue, as a clear table. Include the immediate-escalation overrides.
4. Confidence handling — what happens below a threshold (e.g. confidence < 0.7): route to a human-triage queue, never guess and silently misroute. State the threshold.
5. Priority signals — the cues that bump a ticket up (angry sentiment plus paying customer, churn language, outage keywords, vulnerable-customer signals) and the cues that keep it standard.
6. Failure modes & guardrails — what it must NEVER do: invent a category outside the list, change ticket content, contact the customer, or suppress an escalation.

Rules:
- The agent classifies and routes only. It does not answer the customer.
- It must output only the schema fields. No free-text replies to users.
- Any safety, security, or vulnerable-customer signal is an immediate escalation, overriding normal category logic.
- If a ticket cannot be confidently classified, default to human triage — never force a guess.

Output: the system prompt, the classification schema, the routing table, confidence handling, priority signals, and failure guardrails.

Success signal: the output is good only if the agent returns a fixed schema with a confidence score, low-confidence tickets route to humans, and every safety/security signal is an immediate override escalation.