How to Integrate AI Agent Safety with PagerDuty

SafeClaw by Authensor integrates with PagerDuty to escalate critical AI agent safety events into your incident management workflow. When an AI agent attempts a dangerous action, fails an audit integrity check, or triggers repeated policy violations, SafeClaw creates a PagerDuty incident automatically. SafeClaw supports Claude and OpenAI agents, has 446 tests, and provides hash-chained audit logs.

Prerequisites

• SafeClaw installed (npx @authensor/safeclaw)
• A PagerDuty account with a service configured
• A PagerDuty Events API v2 integration key

Step 1: Create a PagerDuty Service Integration

1. Log into PagerDuty and navigate to Services > Service Directory.
2. Select an existing service or create a new one named "AI Agent Safety".
3. Go to the Integrations tab and click Add Integration.
4. Select Events API v2 and click Add.
5. Copy the Integration Key (also called a routing key).

Step 2: Configure SafeClaw for PagerDuty

Add PagerDuty settings to your .safeclaw/policy.yaml:

version: 1
default: deny

notifications:
  pagerduty:
    routing_key: "${PAGERDUTY_ROUTING_KEY}"
    severity_map:
      action.denied: warning
      audit.integrity_failure: critical
      policy.violation: error
      action.prompted: info
    dedup_key_prefix: "safeclaw"

rules:
  - action: file.read
    paths:
      - "src/**"
    decision: allow

- action: file.write
    paths:
      - "src/**"
    decision: prompt

- action: shell.execute
    decision: deny

- action: network.request
    decision: deny

Set the environment variable:

export PAGERDUTY_ROUTING_KEY="your-integration-key-here"

Step 3: Configure Severity Escalation

Map SafeClaw events to PagerDuty severities for proper routing:

notifications:
  pagerduty:
    routing_key: "${PAGERDUTY_ROUTING_KEY}"
    escalation:
      critical:
        events:
          - audit.integrity_failure
        auto_resolve: false
      high:
        events:
          - action.denied
        threshold:
          count: 5
          window_minutes: 10
        auto_resolve: true
        resolve_after_minutes: 30
      low:
        events:
          - action.prompted
        threshold:
          count: 20
          window_minutes: 60

This configuration creates a critical incident immediately on any audit integrity failure. For denied actions, it waits until 5 denials occur within 10 minutes before escalating. Prompted actions only trigger a low-severity alert after 20 occurrences in an hour.

Step 4: Configure Deduplication

PagerDuty uses deduplication keys to group related events. SafeClaw generates these automatically:

notifications:
  pagerduty:
    routing_key: "${PAGERDUTY_ROUTING_KEY}"
    dedup_key_prefix: "safeclaw"
    dedup_strategy: "action_type"

With action_type deduplication, all shell.execute denials group into a single PagerDuty incident, preventing alert storms while still capturing every individual event in SafeClaw's audit log.

Step 5: Add Custom Details to Incidents

SafeClaw includes rich context in the PagerDuty event payload:

{
  "routing_key": "your-key",
  "event_action": "trigger",
  "dedup_key": "safeclaw-shell.execute-20260213",
  "payload": {
    "summary": "SafeClaw: shell.execute denied (5 occurrences in 10 min)",
    "severity": "high",
    "source": "safeclaw-agent-monitor",
    "component": "ai-agent",
    "custom_details": {
      "action_type": "shell.execute",
      "target": "rm -rf /tmp/data",
      "agent": "gpt-4o",
      "policy_rule": "deny",
      "audit_hash": "a3f2...b71c",
      "recent_log_entries": 5
    }
  }
}

Step 6: Test the Integration

Send a test event to PagerDuty:

npx @authensor/safeclaw test-notify --channel pagerduty

Verify the incident appears in your PagerDuty dashboard. Then trigger a real denied action and check that the incident is created with the correct severity and deduplication.

Summary

SafeClaw integrates with PagerDuty to bring AI agent safety into your incident management workflow. Severity mapping, threshold escalation, and deduplication ensure your on-call team receives actionable alerts without noise. Hash-chained audit logs provide the forensic detail needed during incident investigation. SafeClaw is MIT licensed and open source.

Related Guides

• How to Send AI Agent Safety Alerts to Slack
• How to Track AI Agent Errors in Sentry
• How to Get SMS Alerts for AI Agent Safety Events
• How to Monitor AI Agent Actions in Datadog