AI triage: a summary on every new incident

When an incident opens, Shorthop can hand it to an AI agent of your own. The agent looks into it with whatever tools you give it (logs, metrics, deploys) and writes back what is wrong, the likely cause and how to fix it as a note on the incident. The note appears on the incident page, on every open alert's timeline, in the incident's Slack thread and as a comment on its Jira issue.

Shorthop does not run the agent. It sends the incident out and accepts the answer back, so the agent, its model and its tools stay yours: an n8n workflow, a small service, anything that can receive a webhook.

incident opens ─▶ webhook channel ─▶ your agent ─▶ your tools (logs, metrics, deploys…)
                                         │
Shorthop ◀── note, marked "suggestion" ──┘

1. A token for the agent

Create a user for the agent (for example ai-triage@…) and, under My account → API & MCP, a token with alerts:read and alerts:write. Nothing more: the agent only reads and writes notes. Its name is what the note shows as its author.

2. A webhook channel to the agent

Channels → New → Webhook, pointing at your agent, with events: triggered so it runs once per incident rather than on every escalation and resolve. A body template that gives the agent what it needs:

{
  "incident_id": "{{incident_id}}",
  "title": "{{title}}",
  "severity": "{{severity}}",
  "url": "{{url}}",
  "alerts": "{{alerts_text}}",
  "history": "{{history}}"
}

history is the line Shorthop already knows ("Seen 3 times before. Last fixed by …"). Pass it on: what fixed this alert last time is often the best lead the agent will get.

Then add the channel as a target in the first step of the escalation policies whose incidents should be triaged, next to the people. The agent then starts at the same moment the page goes out, and never delays it.

3. The agent writes back

Over MCP (the agent is usually an MCP client already):

add_incident_note(incident_id, text, suggestion=true)

or over the REST API:

curl -X POST https://oncall.internal.example.com/api/v1/incidents/$INCIDENT_ID/notes \
  -H "Authorization: Bearer sh_…" -H "Content-Type: application/json" \
  -d '{"text": "…", "kind": "suggestion"}'

Why "suggestion"

A note with kind: "suggestion" (or suggestion=true) is shown as suggestion · unverified, apart from people's notes, in the web app, the phone app and Slack. It is a guess written by something that was not there, and a responder should never mistake it for a colleague's finding. It does not feed an alert's history either: "what fixed it" is only ever what a person records when resolving.

Good practice

Edit this page on GitHub