AI Agent Incident Response Plan Template

Copyable incident-response plan

# AI Agent Incident Response Plan

## 1. Incident identity
- Incident ID:
- Start time / detection time:
- Owner:
- Affected agent/workflow:
- Affected provider/account/system:
- Severity: SEV-1 / SEV-2 / SEV-3 / SEV-4

## 2. Known external effect
- Last known good effect:
- First known bad or ambiguous effect:
- Is the harmful effect still propagating? yes/no/unknown
- Money, messages, files, deployments, credentials, or accounts affected:
- Provider-native IDs / URLs / transaction IDs:

## 3. Immediate containment
- Exact action path frozen:
- What remains intentionally online/read-only:
- Automatic retries paused for ambiguous side effects:
- Credentials or sessions revoked/isolated, if necessary:

## 4. Independent evidence captured
- Agent-visible request/context:
- Tool/API request and response:
- Provider-side state:
- Runtime/commit/config identity:
- Screenshot/UI state, when relevant:
- Other independent source:

## 5. Failure classification
- Decision failure?
- Execution/tool failure?
- Stale/freshness failure?
- Wrong target / identity / custody failure?
- Duplicate/idempotency failure?
- Runtime/source divergence?
- Observability/reporting mismatch?
- Evidence supporting the leading hypothesis:
- Evidence that could disprove it:

## 6. Recovery
- Smallest reversible repair:
- Canary or first restored action:
- Provider/runtime/UI evidence that recovery worked:
- Remaining uncertainty:

## 7. Customer / counterparty impact
- Who was affected:
- What needs correction/refund/reversal/notification:
- Exact completion state:

## 8. Recurrence prevention
- Trigger → decision → action → effect → detection → recovery chain:
- Durable code/config/process change:
- What should be deleted or simplified:
- What signal would catch recurrence earlier:
- Owner and completion evidence:

## 9. Closure
- External effects reconciled:
- Recovery independently verified:
- Follow-up work separated from incident closure:
- Closure time: