Agent support desk

Agent support desk

Point one agent at a ticket queue and it guesses severity, answers before it has reproduced anything, and escalates either everything or nothing. A desk divides the labor: one agent sorts, one reproduces, one writes, one decides when a person takes over. The shape is a sequential pipeline with a capped repro loop and an escalation branch that always exits to a human.

Method

1. Triage agent sorts and dedupes. Load support-engineer-role. Read the inbound ticket and emit ticket.yaml: severity S1 to S4, product area, customer tier, and a dedupe key tying it to any open incident. An S1 outage bypasses the queue and drops straight into the escalation branch.
2. Repro agent isolates the failure. Reproduce the reported behavior in a clean environment and write repro.md: minimal steps, exact versions, actual against expected. When the ticket lacks the detail to reproduce, reply with a named list of missing facts rather than guessing at the cause.
3. Cap the repro loop at two rounds. After two requests for information, mark the ticket cannot-reproduce and send it to escalation carrying what was tried. A ticket that spins forever on missing data starves the rest of the queue.
4. Drafter writes the customer reply. Turn repro.md into reply.md: acknowledge, state the cause in the customer's words, give the fix or a workaround, name the next checkpoint. Hold the product's voice and let no internal jargon reach the customer.
5. Escalation agent owns the handoff. Route to a person when severity is S1 or S2, the customer is enterprise tier, the fix needs a code change, or repro failed. Write escalation.md: summary, repro state, suggested owner. Everything else stays inside the pipeline.
6. A human approves before send. No agent sends a customer-facing message. The drafter's reply.md waits for a person to approve or edit it, because the send is the human's call, not the desk's.

Run it

In Claude Code, run the four roles as subagents behind an orchestrator that reads ticket.yaml to choose the route and passes each artifact as a file on the ticket's branch. Terminate when a human approves and sends the reply, or when the ticket reaches a person through the escalation branch. To port, model it as a CrewAI sequential Crew with a conditional escalation task, an AutoGen GroupChat whose human proxy agent approves the reply, or a LangGraph graph with a router node and a human-in-the-loop interrupt before send.

Signals it works

• Every reply a customer received was approved by a person first.
• Triage severity matches what the ticket turned out to be once resolved.
• Cannot-reproduce tickets escalate with the steps tried, not a blank field.

Boundaries

The desk handles inbound triage and drafting, not the fix: a code change hands off to the engineering team's own flow. It does not set severity policy or refund authority, which follow company convention, and human judgment owns anything irreversible, from sending a reply to closing an enterprise escalation.