Support triage where the AI drafts and humans keep the veto

Trigger: helpdesk webhook · Pattern: classify → confidence gate → send or pause · Sample: examples/use-cases/support-triage.toml · Status: runs today (intel-remote,schema,tools-http-tls)

The problem

Support teams don't drown in hard tickets; they drown in the hundred easy ones between each hard one. "How do I reset my API key" deserves a fast, correct, warm answer — it does not deserve twenty minutes of a senior engineer's queue time. But full auto-reply is how you get the other failure: a confidently wrong bot answer on a P1, screenshotted, viral.

The question every support lead actually asks: can the AI answer the easy ones and reliably know which ones aren't easy?

What the agent does

1. The helpdesk webhooks every new ticket to the workflow.
2. One schema-enforced LLM step does three jobs at once: classifies severity (p1|p2|p3) and topic, drafts a reply in your voice, and — the load-bearing part — rates its own confidence as high or low. The schema makes that a forced binary, not a hedge.
3. A switch on confidence routes the graph:
   • high → the reply posts to the ticket immediately.
   • low → the run checkpoints and stops (pause_for_approval). A human reads the draft — in the helpdesk, or via agentd inspect — edits if needed, and resumes the run to send.
4. Every classification, draft, and decision lands in the audit log.

The confidence gate is the product

Two design choices make this deployable where naive auto-reply isn't:

• The gate is a declared edge, not a vibe. "When confidence is low, a human approves" is written in the workflow file — reviewable in a pull request, signable, and impossible for the model to route around. The model can be wrong about its confidence; it cannot act on that wrongness beyond producing a draft a human reads.
• The pause is durable. A paused run is a checkpoint on disk with a run id, not a thread blocking in memory. Restart the daemon, come back Monday — --resume RUN_ID picks up exactly where it stopped. Human-in-the-loop that survives a deploy.

[[edges]]
from = "gate"
when = "high"
to = "reply_url"        # straight to the customer

[[edges]]
from = "gate"
when = "low"
to = "human_review"     # checkpoint; a person owns the send

Turning the dial as trust grows

Day one, you might route everything through human_review — the agent as a drafting assistant. After a month of audit logs showing the high drafts were shippable, flip the high edge to post directly. After a quarter, maybe p3 + high skips review and p1 never does. Each step is a one-line edge change with a paper trail — and the conformance suite can hold a regression bar over the classifier (min_pass_rate) so a model update that degrades triage fails CI instead of failing customers.

Honest limits

• Replies post via the helpdesk's HTTPS API (tools-http-tls). If your helpdesk only does email-out, pair this with the inbox concierge pattern instead.
• Retrieval is the next increment: wiring a knowledge-base MCP server (read_mcp_resource) into the draft step grounds answers in your docs. The pattern doesn't change — one more bounded read before the same gate.