{
  "schema": "varexis-public-evidence-record/0.3",
  "generated_at_utc": "2026-10-09T18:26:37.372Z",
  "record_id": "VXI-EXT-OAI-AGENTS-GR-03-v1",
  "record_kind": "external_reference",
  "title": "OpenAI Agents SDK: Where guardrails apply",
  "system": "OpenAI Agents SDK",
  "evidence_family": "agent_tool_validation_coverage",
  "index_status": "PUBLIC_SAFE_SOURCE_REVIEW",
  "evidence_state": "PRIMARY_DOCUMENTATION_REVIEW_ONLY",
  "canonical_question": "Where should validation run when an agent hands work to another agent or calls a tool?",
  "near_questions": [
    "Do input guardrails check every agent after a handoff?",
    "Which agent runs the final output guardrail?",
    "Where should validation check tool arguments?",
    "Do agent guardrails cover every intermediate tool call?",
    "Can a manager agent rely only on its entry checks?",
    "How should guardrail coverage be reviewed before enabling tools?",
    "Are hosted tools covered by attached function-tool guardrails?",
    "What validation belongs beside a side effect?",
    "Does configuring a guardrail demonstrate that it ran?",
    "Which tool types need separate control checks?",
    "What should a handoff safety review record?",
    "Can final answer validation prevent an earlier tool action?"
  ],
  "supported_claims": [
    {
      "claim": "OpenAI documents input guardrails on the first agent, output guardrails on the final-output agent, and tool guardrails on the function tools to which they are attached. Review validation at each side-effecting tool boundary; entry and exit checks alone do not cover every intermediate action. This documents control placement, not tested enforcement or protection for every tool type.",
      "reliance": "DOCUMENTED_CAPABILITY_ONLY"
    }
  ],
  "not_demonstrated_claims": [
    "automatic validation of every intermediate action",
    "uniform guardrail coverage across all tool types",
    "tested enforcement in a buyer deployment"
  ],
  "public_summary": "OpenAI documents input guardrails on the first agent, output guardrails on the final-output agent, and tool guardrails on the function tools to which they are attached. Review validation at each side-effecting tool boundary; entry and exit checks alone do not cover every intermediate action. This documents control placement, not tested enforcement or protection for every tool type.",
  "related_record_ids": [
    "VXI-EXT-OPENAI-AGENTS-SDK-02-v1",
    "VXI-EXT-OAI-AGENTS-TRACE-01-v1"
  ],
  "source_review": {
    "source_type": "official_external_documentation",
    "source_title": "OpenAI Agents SDK: Guardrails and human review",
    "source_url": "https://developers.openai.com/api/docs/guides/agents/guardrails-approvals",
    "checked_on": "2026-10-09",
    "candidate_record_id": "VXI-EXT-OAI-AGENTS-GR-03-v1",
    "candidate_record_sha256": "7ff90a05865275e2c7de806d9c21620e879b214c1d4e857db85a6b996c09c302",
    "method": "Primary documentation reviewed for the published bounded statement.",
    "deployment_tested_by_varexis": false,
    "revalidation": "The delivered snapshot preserves the review date shown. Recheck current primary documentation and deployment context before relying on present applicability; payment does not trigger fresh research."
  },
  "answer_policy": {
    "classification": "DOCUMENTED_WITH_DEPLOYMENT_BOUNDARY",
    "bounded_answer": "OpenAI documents input guardrails on the first agent, output guardrails on the final-output agent, and tool guardrails on the function tools to which they are attached. Review validation at each side-effecting tool boundary; entry and exit checks alone do not cover every intermediate action. This documents control placement, not tested enforcement or protection for every tool type.",
    "required_buyer_context": "SDK language and version, agent and handoff graph, tool types, attached checks and observed execution results. This is control-placement guidance, with no security test.",
    "missing_evidence": "A broader claim needs evidence matching the intended use. Payment does not upgrade the evidence state."
  },
  "private_boundary": "Raw logs, thresholds, private paths, source code, protected mechanisms and reconstruction-level details are excluded from public and paid delivery.",
  "paid_access_note": "The full published answer and boundaries are free. Paid access saves the selected route and source versions, with an unsigned receipt when purchased. Preview coverage before paying. Fresh research, bespoke assessment and protected evidence are excluded.",
  "service_route": {
    "request_guide": "/varexis-index/request.html",
    "request_workflow": "/varexis-index/request-workflow.json",
    "service_id": "single_request",
    "service_routes_url": "/varexis-index/service-routes.json",
    "payment_required": true,
    "availability_endpoint": "/api/index/capabilities",
    "operational_status": "AUTOMATIC_SERVICE_CAPABILITIES_GATED",
    "checkout_url": "/index-access"
  },
  "authority_boundary": "VAREXIS-INDEX informs evidence reliance. It does not provide advice, certification, deployment authority, security assurance, production approval, or guarantee.",
  "human_page": "/varexis-index/catalog.html#agent-tool-guardrail-coverage",
  "machine_readable_record": "/varexis-index/records/agent-tool-guardrail-coverage.json"
}
