Is action_request_id a field from the embedded audit event?

No. action_request_id is added for standalone publication so the request record has its own addressable identity. The embedded object is named audit_event.requested_action.

Does this record contain the action arguments?

No. The record identifies the requested capability, purpose, target, boundary, environment, and optional financial amount. Runtime producers keep argument material in their own controlled records and can link hashes or evidence separately.

Can a request be used without a policy decision?

It can be recorded before policy evaluation. A governed execution uses the correlation object to connect the request to the resulting policy decision.

Technical reference · v1.0.0

AI Agent Action Request Schema

An action request records the authority an AI agent asks to use before policy evaluation and any tool effect. This standalone reference preserves the requested_action shape from the published audit event and adds an addressable action_request_id plus correlation for independent use.

Download JSON Schema Download minimal example

JSON Schema draft 2020-12 · Version 1.0.0 · Normative field names in English

Definition: A standalone record of one requested agent action: capability, purpose, resource, data boundary, environment, and request time.
When used: Create it when an agent proposes a governed action. Link it to policy, approval, tool, and audit records with correlation.
Required fields: The request ID, schema version, correlation IDs, action, purpose, resource, data boundary, environment, and requested_at.
Optional detail: Include amount for financial actions and trace context when the execution carries OpenTelemetry identifiers.

Schema and examples

The versioned artifacts are self-contained. Consumers can validate the request without loading the audit-event schema.

JSON SchemaDraft 2020-12 JSON Schema, version 1.0.0./ai-agent-action-request/v1/schema.jsonDownload Basic read exampleMinimal synthetic request with no financial amount./ai-agent-action-request/v1/examples/basic-read.jsonDownload Credit disposition exampleSynthetic financial action with amount and trace correlation./ai-agent-action-request/v1/examples/credit-disposition.jsonDownload

Object and execution order

The request is the first durable contract in the reference sequence. Policy evaluates it before an approval or tool effect is recorded.

01RequestThe agent proposes an action against a named resource inside a declared data boundary and environment.
02CorrelateThe action_request_id identifies this standalone record. correlation_id and execution_id connect it to the execution graph.
03EvaluateThe policy decision records allow, warn, require_approval, or block for the same execution correlation.
04ContinueA tool call follows an allowed or warned route. A require_approval route records an approval event before an effect.

Field dictionary

The dictionary covers every top-level field, nested resource member, correlation member, and conditional financial member.

Standalone identity and correlation

Field	Status	Purpose
schema_version	Required	Selects the compatibility contract used to parse the record.
action_request_id	Required	Independently addresses this request and links it to an audit event. Added for standalone publication.
correlation.correlation_id / execution_id	Required	Joins the request to policy, approval, tool, and audit records without a timestamp join. Added for standalone publication.
correlation.trace_id / span_id / parent_event_id	Optional	Carries OpenTelemetry trace context and an optional parent event. All-zero W3C identifiers are invalid.

Requested authority

Field	Status	Purpose
action	Required	Names the capability the agent proposes to use.
purpose	Required	States the business purpose supplied to policy evaluation.
resource.type / resource.id	Required	Identifies the target resource and its resource kind.
data_boundary_ref	Required	Names the governed data boundary associated with the request.
environment	Required	Identifies the execution environment supplied to the request.
requested_at	Required	Records when the action was requested as an RFC 3339 date-time.
amount.value / amount.currency	Conditional	Carries a decimal string and three-letter uppercase currency when financial value affects policy.

Minimal example

This example contains the required fields for a synthetic read action. The second downloadable example adds amount and trace context.

{
  "schema_version": "1.0.0",
  "action_request_id": "req_action_demo_0001",
  "correlation": {
    "correlation_id": "corr_credit_review_demo_0001",
    "execution_id": "exec_credit_review_demo_0001"
  },
  "action": "read_case_summary",
  "purpose": "review_credit_case",
  "resource": {
    "type": "credit_case",
    "id": "case_demo_0001"
  },
  "data_boundary_ref": "boundary_credit_case_minimum",
  "environment": "staging",
  "requested_at": "2026-07-21T09:14:29.100Z"
}

Validation and digest verification

JSON Schema validation checks shape. The companion verifier reproduces a sha256: digest over the canonical record and can check an optional detached Ed25519 signature.

01Load the versioned schema from the JSON Schema download URL.
02Validate the JSON document with a draft 2020-12 validator and a date-time format plugin.
03Canonicalize the complete standalone record with recursively sorted object keys, matching the published audit-event verifier approach.
04Compute SHA-256 over the canonical bytes and represent the result with the sha256: prefix.
05Compare the computed digest with the digest retained by the calling evidence procedure.
06When a detached Ed25519 signature is supplied, resolve its public key independently and verify the signature over the 32-byte digest.
07Record a hash mismatch or signature failure as a failed verification result and preserve the original record for investigation.

Schema-valid request

The downloadable examples validate against draft 2020-12 and produce a reproducible sha256: digest from their canonical content.

Tamper check

Changing a field such as action keeps the document structurally readable while changing the canonical digest. The verifier reportsrecord_hash_mismatch.

The standalone verifier accepts an optional detached Ed25519 signature. Trust in a public key comes from the caller’s independently governed key registry.

Compatibility and versioning

Version the standalone contract independently from agent, policy, and tool releases.

Patch

Clarifications, descriptions, and examples can change while validation behavior remains stable. A corrected immutable artifact receives a new versioned URL.

Minor

New optional fields require a new schema_version and versioned URL. Consumers declare support before processing the new version.

Major

Removed or renamed fields, changed meanings, stricter required status, or changed canonicalization require a new major path.

Consumers preserve unknown versions for review and stop semantic processing until support is declared. The v1 path uses schema_version 1.0.0.

How KLA implements this

The embedded audit schema is the normative field source. Runtime producers expose related fields that the public standalone shape can link and normalize.

Action request fields mapped to current KLA implementation sources
Contract area	Current source	Mapping	Status
Embedded requested action shape	apps/platform/src/lib/ai-agent-audit-event/v1/schema.json	The audit-event schema defines requested_action.action, purpose, resource, data_boundary_ref, environment, amount, and requested_at. This page publishes those fields at the top level.	Shipped reference source
Runtime tool request facts	services/execution-worker/src/services/audit-events.ts	recordToolAuditEvent records executionId, toolName, toolAction, data scopes, and hashes for tool parameters and results. The function does not emit this normalized standalone request object.	Current related producer
Execution correlation	services/execution-worker/src/services/audit-events.ts	Tool audit records carry executionId and the active OpenTelemetry trace ID. Consumers use the standalone correlation object to join those runtime facts.	Current related producer
Policy and audit linkage	services/shared/src/policy/contracts.ts	GateContext and GateDecision contracts provide the policy-side execution identifiers and canonical decision values used after this request is evaluated.	Current consumer contract

Deliberate abstractions and unsupported fields

•The schema omits tenant database IDs, internal row IDs, and Cerbos policy syntax.
•The schema carries resource references and data_boundary_ref. It does not carry raw resource contents, prompts, model output, tool arguments, or tool results.
•The schema does not assert that the requester is authorized. The policy decision and the consuming authorization path establish the applicable control result.
•The schema does not prescribe how an implementation resolves a data boundary reference or retains the referenced data.
•The standalone action_request_id and correlation object are additions for standalone publication. The embedded requested_action object has neither field.
•The public examples contain synthetic IDs and timestamps. They carry no customer data or credential material.

Related references

Follow the governed execution sequence from request through policy, approval, and the full audit event.

Policy decision schema

The policy result evaluated for this requested action.

Approval event schema

The maker-checker record used when policy requires human approval.

Audit log schema

The full event envelope that connects request, policy, approval, tools, and outcome.

Apply the contract

Connect the request to the rest of the governed execution record.

Use the audit log schema for the complete event envelope, then follow the linked policy and approval references through the execution sequence.

AI agent audit log schema AI agent access control