Build an Incident-Triage Agent for Operators

From the AgentsBooks dashboard click **+ New Agent**. Pick the **Custom Agent** preset on the wizard's first card, then on step two enter:

• Name: Halt
• Role: On-call Triage

Halt is our worked example. The playbook teaches you how to build an incident-triage agent, and we use a one-syllable name because half the time you're naming it from your phone at 2 AM. Click ✨ Create Agent. The empty profile hub opens — eight cards waiting to be wired.

Open the **Personal** card. Triage agents fail two ways — too noisy, or too calm during a real fire. Personality keeps Halt on the right side of that line. Set:

• Traits: calm under pressure, precise, evidence-driven
• Communication style: terse incident-report cadence with cited timestamps
• Tone (default): concise operational
• Voice ID: halt-grave · Provider: elevenlabs · Pace: measured · Pitch: low

Three traits is the sweet spot. The voice block matters because Halt narrates the daily 5 PM summary in the team Slack huddle.

Open **Brain**. Pick a low-temperature reasoning model — we use `claude-sonnet-4-6` at temperature `0.2` — and paste the four-rule system prompt:

You are Halt, an incident triage agent. Always classify severity using the
runbook matrix — never guess. Open exactly one Slack thread per incident;
cross-link related events from long-term memory. Page the on-call human
only for sev-1 or repeated sev-2. Refuse to auto-resolve any incident
that affects authentication, billing, or data integrity.

Low temperature matters here. You want Halt to follow the matrix, not improvise. The four rules are a contract — every behaviour in the loop maps back to one of them.

Open **Knowledge** and add three texts plus two URLs. Halt retrieves from this on every classification call — this is what keeps the page rate low. Upload at minimum:

• A per-service runbook (auth, billing, data-pipeline, web-edge — one section each)
• A severity matrix with explicit sev-1, sev-2, sev-3 rules
• The last 25 post-mortems so Halt cites prior precedent
• URL: your status page archive (hourly refresh)
• URL: the internal runbook docs (daily refresh)

No runbook yet? Paste a one-page placeholder with the three severity tiers and one detection signal per service. Halt will work with that and ask clarifying questions on the first sev-1.

Open **Memory** and add a long-term store:

• Name: incident-history
• Type: vector_db
• Default: ✅ on
• Purpose (in config): Per-service incident archive with severity, root cause, time-to-resolve. Source of pattern detection across alerts.

Memory is the difference between triage and noise. Knowledge is what Halt knows about the matrix; memory is what he remembers about your specific systems. Combined with the daily 5 PM summary task, this is what surfaces repeat offenders before they become a sev-1.

Open **Heart** and create one task with **two triggers**:

• Name: Yesterday's incident summary
• Trigger 1: Schedule · Cron 0 17 * * * · Timezone America/New_York
• Trigger 2: Webhook · Path /incidents/inbound
• Prompt: On schedule: pull last 24h of incidents from incident-history, summarise sev-1+sev-2 with cause and resolution, post to #incidents Slack. On webhook: classify severity from payload, open a Slack thread, write to incident-history, page on-call if sev-1.
• Memory namespace: incident-history · Read+Write: ✅

Dual triggers are the move here. The schedule is the daily reflective summary; the webhook is the real-time loop. One task, two doors in.

All seven cards are wired. Open Halt's profile hub — every section now shows a green check. Hit **Publish**. What you have:

• Webhook intake at /incidents/inbound — your alerting stack POSTs every event here, Halt classifies in seconds.
• One Slack thread per incident in #incidents, with related events cross-linked from incident-history.
• Daily 5 PM summary of the last 24 hours, surfacing sev-1 and sev-2 with cause and resolution.
• Auth, billing, and data-integrity events routed to humans only — Halt refuses to auto-resolve them.
• A starting point you can clone with the button on this playbook page — your triage agent in two clicks instead of seven.