Agent Decisions

Use Agent Decisions to send the correct outcome for content that SafetyKit has reviewed. These decisions help SafetyKit measure review quality and improve automated decisions over time.

Use the Agent Decisions API reference for the full request schema, SDK examples, and response fields.

When to send decisions

Send an agent decision after a human reviewer, internal agent, or downstream system determines the correct outcome for a SafetyKit review.

Each request should represent one content object and one policy category. If the same object is reviewed against multiple policy categories, send one request per policy category.

Request fields

Field	Type	Description
reference_id	string	Stable identifier for the content object in your system, such as a listing ID, post ID, campaign ID, or account ID.
content_type	string	SafetyKit data namespace or content type for the content being reviewed.
policy_id	string	SafetyKit policy category ID that was evaluated, as configured for your account.
decision	string	Human reviewer’s verdict for the policy category, such as "match" or "no_match".
agent_type	string	Optional type of review that produced this decision, such as "appeal", "bpo_review", or "policy_expert".
action	string	Optional correct enforcement action for this content, such as "reduce_visibility" or "unpublish".
comment	string	Optional additional context or details from the reviewer.

Sample request

curl -X POST "https://api.safetykit.com/v1/agent_decisions" \
  -H "Authorization: Bearer <SAFETYKIT_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "reference_id": "listing_12345",
    "content_type": "product_listing",
    "policy_id": "counterfeit",
    "decision": "no_match",
    "agent_type": "appeal",
    "comment": "The listing is allowed under the counterfeit goods policy."
  }'

Response:

{
  "safetykitCorrect": true
}

Best practices

• Use stable reference_id values that match the IDs you send to SafetyKit during ingestion.
• Keep content_type consistent with the namespace or object type SafetyKit reviewed.
• Send the final corrected decision, not an intermediate review state.
• Retry transient network errors and 5xx responses with exponential backoff.
• Do not retry 4xx responses without changing the request.