Skip to Content
ExamplesTriage support tickets

🎧 Triage support tickets and draft replies

Stack — Trigger: agent API / Slack · Models: GPT-4o primary + GPT-4o & Gemini 1.5 Pro verifier panel · Tools: order-lookup · Mode: verify + approval · Memory: on

Who it’s for — support teams drowning in L1 tickets who can’t let an agent send a wrong answer.

This recipe is a real, running agent — the Customer Support Orchestrator below has handled 8 runs at 87.5% success for ~$0.07 total. Everything shown is its actual configuration and history.

The Customer Support Orchestrator: primary GPT-4o plus a cross-vendor verifier panel

What it actually does

It reads a ticket, looks up the order, drafts a reply, and a cross-vendor verifier panel checks the draft before it ships. Two real runs from its history:

  • “Where is my order NW-10428?” → it called the Lookup Order Status tool, and the verifier caught a delivery-date inconsistency (the ETA didn’t match the carrier’s speed) and forced 1 revision before the reply passed. 3 steps, 6,463 tokens, $0.0121.
  • “URGENT ticket TKT-6190 from Brightway Logistics (enterprise customer)…” → it hit the refund-escalation rule and stopped at waiting approval for a human, instead of promising something it shouldn’t.

Real runs — a completed order-status reply, an enterprise ticket paused at waiting-approval, a failed run

Build it

  1. Create the agent (Agents → Create → Autonomous) with a primary model — this one uses GPT-4o at temp 0.3 — and instructions to triage the ticket, look up the order, and draft a reply in your voice.

  2. Attach the tools it needs — here an order-lookup tool generated from a connected orders API. (Connect yours in APIs → Connect API.)

  3. Add a verifier panel (Models & Verification → Configure): add reviewers from more than one vendor — this agent runs GPT-4o + Gemini 1.5 Pro with any_fail_blocks and up to 2 revisions, triggered on_final_output. The primary drafts, the reviewers try to refute, and it revises before anything ships.

  4. Set constraints — always-on rules it can’t argue past. This agent’s include “Never promise a refund over $500 without escalating to a human first” and “Do not invent order numbers or account details that are not in the ticket” — plus one it learned from a failed run (“Avoid the failure encountered previously: LLM provider is not healthy”).

    The agent's four constraints, including a learned one

  5. Turn on memory so it carries context between runs, and gate refunds with request_approval — that’s what paused the Brightway Logistics ticket above.

  6. Wire the trigger — call the agent’s API when a ticket arrives (its endpoint is on the Overview → Integration panel), or publish it on Slack.

Make it yours

  • Swap the orders API for Zendesk/Intercom — same Connect API step.
  • Tighten or loosen the refund threshold in the constraint.
  • Promote a good run to a skill (button on any completed run) so other agents reuse the flow.

Uses — Support · Verification · Approvals · Memory

OverviewDeflect L1 from your docs