A support triage agent for NeIO Meeting Assistant. Per ticket it makes six tool calls: read the record, apply the privacy gate, look for duplicates, decide a route, verify the decision, then act. Press Run and watch it work the queue.
Source: the bundled sample data/tickets.json, 25 tickets across 5 accounts and 6 areas. ZeAlert Bug #32872 was not reachable from this session, so the sample stands in, as the brief allows. This page replays deterministic rules: no model calls at runtime, and every action stays a draft. Nothing is sent.
Nothing yet. Run plays the whole queue; Step advances one tool call.
0 + 0 + 0 + 0 = 0 processed
Escalations land here and stop. The agent does not touch privacy requests.
On-call pages and outbound replies also wait for a person before send. This demo never sends.
read_ticket(id) → the full recordcheck_privacy_policy(area, body) → requires_human, true or falsefind_duplicates(subject, seen) → earlier tickets with the same subjectdecide_route(priority, policy, dups) → escalate, page, route, or linkverify_decision(decision, ticket) → asserts before acting; any failure forces escalateact(...) → approval card, page draft, queue assignment, or link. Drafts only.The honest tradeoff. Run live, this agent spends an estimated 4,000 tokens per ticket (six tool calls, context replayed each round) against roughly 200 per ticket for a single prompt that classifies the whole batch. That is about 20x. What it buys: a verify step before every action, and a human gate in front of all 4 privacy deletions, every P1 page, and every outbound reply.
Built with Claude Code in The Lab · Yashasvi Shukla · Codvo.AI · 07-Sep-2026 · Data: data/tickets.json (bundled sample)