Ticket Triage Agent — NeIO Meeting Assistant

A step-by-step walkthrough of a triage agent design: what tool it calls, what it decides, what it checks before acting, and where a human has to sign off.
Data source: User Story #33669 (TFE – Catalogue Revamp to Agentic + Cataloging Agent) was the intended source, but no Azure DevOps connection is available in this environment — so this walkthrough runs on the bundled sample instead: data/tickets.json, 25 support tickets for NeIO Meeting Assistant.

Fields missing or malformed are handled with safe defaults. Inputs over 150 tickets switch to fast/instant rendering automatically.

0
Processed
0
Auto-resolved
0
Escalated · privacy
0
Escalated · pattern
Press Run to step through the tickets.
Cost tradeoff: this design runs 3–6 tool calls per ticket (fetch → parse signals → check human flag → check pattern → match playbook → draft). At roughly 120–250 tokens per call, that's an estimated 600–1,500 tokens/ticket — call it 5–8× the cost of one batch prompt that classifies all 25 tickets in a single call (~3–5k tokens total). What the extra spend buys: a per-ticket audit trail and a hard stop before anything irreversible — of these 25 tickets, 4 privacy/deletion requests and 1 repeating P1 pause for a human; the other 20 draft and log without ever sending.
This page simulates the agent's decision logic with deterministic rules so the design can be inspected step by step — it does not call a live LLM.