Ticket Triage Agent

The tradeoff: this pipeline runs ~5-6 tool calls per ticket (fetch → classify → playbook lookup → gate check → draft → log), each carrying its own context overhead — roughly 8-12× the tokens of triaging the same batch in a single prompt (all 25 tickets in one pass, one output table). What that buys: a hard, auditable gate — step 4 — before anything involving needs_human=true, a privacy-area request, or a P1 is allowed to auto-close. The single-prompt version has no equivalent checkpoint; it just returns an answer. The gate is where a human sits between the agent and the account, every time the ticket is a privacy request, already flagged, or top priority.