How to Track AI Agent Errors in Sentry

SafeClaw by Authensor integrates with Sentry to capture AI agent safety errors as trackable events. When an agent action is denied, a policy violation occurs, or the audit hash chain fails integrity checks, SafeClaw sends a structured error to Sentry with full context. SafeClaw supports Claude and OpenAI agents, has 446 tests, and provides hash-chained audit logging.

Prerequisites

• SafeClaw installed (npx @authensor/safeclaw)
• A Sentry account and project
• A Sentry DSN (Data Source Name)

Step 1: Install SafeClaw

Initialize SafeClaw in your project:

npx @authensor/safeclaw

Step 2: Configure Sentry Integration

Add Sentry settings to .safeclaw/policy.yaml:

version: 1
default: deny

notifications:
  sentry:
    dsn: "${SENTRY_DSN}"
    environment: "production"
    release: "1.0.0"
    events:
      - action.denied
      - audit.integrity_failure
      - policy.parse_error
      - agent.timeout
    tags:
      team: "platform"
      component: "ai-safety"

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

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

- action: shell.execute
    decision: deny

Set the environment variable:

export SENTRY_DSN="https://examplePublicKey@o0.ingest.sentry.io/0"

Step 3: Understand Sentry Event Structure

SafeClaw sends structured events to Sentry with these components:

Error Level Mapping:

• action.denied maps to Sentry warning


• audit.integrity_failure maps to Sentry fatal


• policy.parse_error maps to Sentry error


• agent.timeout maps to Sentry error

{
  "exception": {
    "values": [
      {
        "type": "SafeClawActionDenied",
        "value": "shell.execute denied by policy rule: default:deny"
      }
    ]
  },
  "tags": {
    "action_type": "shell.execute",
    "decision": "denied",
    "agent": "gpt-4o",
    "team": "platform"
  },
  "extra": {
    "target": "rm -rf /tmp/data",
    "policy_rule": "default:deny",
    "audit_hash": "a3f2...b71c",
    "chain_position": 172,
    "recent_actions": ["file.read:allowed", "file.write:prompted", "shell.execute:denied"]
  },
  "breadcrumbs": [
    {
      "timestamp": "2026-02-13T14:31:58.000Z",
      "category": "safeclaw",
      "message": "file.read allowed: src/main.js",
      "level": "info"
    },
    {
      "timestamp": "2026-02-13T14:32:00.000Z",
      "category": "safeclaw",
      "message": "file.write prompted: src/utils.js",
      "level": "warning"
    }
  ]
}

Breadcrumbs show the sequence of agent actions leading up to the error, providing crucial debugging context.

Step 4: Configure Issue Grouping

SafeClaw sets Sentry fingerprints to group related events:

notifications:
  sentry:
    dsn: "${SENTRY_DSN}"
    fingerprint:
      strategy: "action_and_rule"

With action_and_rule grouping, all shell.execute denials by the same policy rule are grouped into a single Sentry issue. This prevents thousands of individual issues while still tracking every occurrence.

Available strategies:

• action_type -- groups by action type only


• action_and_rule -- groups by action type and policy rule


• action_and_target -- groups by action type and target resource


• unique -- every event creates a new issue

Step 5: Set Up Sentry Alerts

In Sentry, go to Alerts > Create Alert Rule:

• When: A new issue is created
• If: Issue is tagged with decision:denied and action_type:shell.execute
• Then: Send a notification to Slack channel #ai-agent-alerts

• When: An event occurs with level fatal
• If: Event message contains "integrity_failure"
• Then: Send to PagerDuty

Step 6: Test the Integration

Send a test event to Sentry:

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

Verify the event appears in your Sentry project dashboard. Then trigger a real denied action:

npx @authensor/safeclaw wrap -- node my-agent.js

Check Sentry for the new issue with full context, breadcrumbs, and tags.

Summary

SafeClaw integrates with Sentry to bring AI agent safety into your error tracking workflow. Structured events with breadcrumbs, tags, and fingerprinting make it easy to investigate and triage agent safety incidents. Sentry's alerting and issue management capabilities complement SafeClaw's deny-by-default approach. SafeClaw is MIT licensed and open source.

Related Guides

• How to Monitor AI Agent Actions in Datadog
• How to Integrate AI Agent Safety with PagerDuty
• How to Send AI Agent Audit Logs to Splunk
• How to Add AI Agent Safety to VS Code