View a markdown version of this page

AgentCore generated Policy in AgentCore observability data - Amazon Bedrock AgentCore
Provided metric dataProvided span data

AgentCore generated Policy in AgentCore observability data

For policy and policy engine resource types, Amazon Bedrock AgentCore publishes invocation metrics to CloudWatch by default. Additional span data is available when traces are enabled for the attached AgentCore Gateway resource, which will emit spans for Policy in AgentCore related operations. See Enabling observability for AgentCore runtime, memory, gateway, built-in tools, and identity resources to learn more about enablement.

Provided metric data

Amazon Bedrock AgentCore publishes the following invocation metrics by default to the AWS/Bedrock-AgentCore CloudWatch namespace. These metrics can be used to observe and monitor policy evaluations and overall performance.

Metric Description Unit

Invocations

Number of requests made to the service

Count

SystemErrors

Number of server-side errors (5xx)

Count

UserErrors

Number of client-side errors (4xx)

Count

Latency

Total time elapsed from sending a request to receiving a response

Milliseconds

AllowDecisions

Number of decisions that resulted in ALLOW

Count

DenyDecisions

Number of decisions that resulted in DENY

Count

TotalMismatchedPolicies

Number of failed policies for a given request due to either missing attribute or type mismatch

Count

PolicyMismatch

Number of failures for a specific policy caused by missing attribute or type mismatch

Count

MismatchErrors

Number of requests that failed due to at least one mismatched policy

Count

DeterminingPolicies

Number of determining policies for a request

Count

NoDeterminingPolicies

Number of requests denied due to no determining policies

Count

GuardrailLatency

Time spent evaluating guardrails for a request

Milliseconds

ConfidenceScore

Score returned by Bedrock Guardrails for a policy evaluation. In the Bedrock Guardrails documentation this value is called a severity score for content filters and prompt attacks, and a confidence score for sensitive information filters. See Bedrock Guardrails score definitions.

None

ConfidenceThreshold

Score threshold configured by a guardrail policy

None

SuppressOutputs

Number of outputs suppressed by policies using the suppressOutput effect

Count

LogOnlyMatches

Number of LOG_ONLY policy matches

Count

LogOnlyDecisionFlips

Number of LOG_ONLY policy matches that change the authorization decision when the policy is promoted to ACTIVE

Count

LogOnlyEvalIncomplete

Number of requests for which LOG_ONLY policy evaluation was incomplete

Count

TemporalLatency

Time spent evaluating temporal policies. One sample is emitted for each temporal evaluation. Use SampleCount to count evaluations. See CloudWatch statistics definitions.

Milliseconds

Metric Dimensions

The following dimensions are available for the above metrics. These dimensions allow you to filter and analyze metric data at finer levels of detail.

Dimension Description

OperationName

The name of the API operation, valid values are AuthorizeAction and PartiallyAuthorizeActions

PolicyEngine

The Policy Engine identifier associated with the metric

Policy

The Policy identifier associated with the metric

TargetResource

The AgentCore Gateway resource identifier associated with the request

ToolName

The name of the tool the metric applies to

Mode

The enforcement mode configured on the AgentCore Gateway, valid values are LOG_ONLY and ENFORCE

Category

The guardrail safeguard type. Valid values are contentFilter, promptAttack, and sensitiveInformation. See Guardrails in policies.

Filter

The guardrail filter or sensitive information entity, such as VIOLENCE, PROMPT_INJECTION, or EMAIL

PolicyEnforcementMode

The enforcement mode of an individual policy. Valid values are ACTIVE and LOG_ONLY. This dimension applies to ConfidenceScore, ConfidenceThreshold, and MismatchErrors. See Test a policy in LOG_ONLY mode.

Provided span data

Amazon Bedrock AgentCore provides additional structured span data through AgentCore Gateway observability, offering deeper insights into API invocations. Policy in AgentCore span data is available after enabling traces for your AgentCore Gateway resource and can be found in CloudWatch aws/spans log group.

Operation Span Attribute Description

AuthorizeAction

aws.agentcore.policy.authorization_decision

The authorization decision after evaluating policies, valid values are ALLOW and DENY

aws.agentcore.policy.authorization_reason

Reason for the authorization decision

aws.agentcore.policy.determining_policies

List of Policy identifiers that determined the decision outcome

aws.agentcore.policy.mismatched_policies

List of Policy identifiers that failed due to missing attributes or type mismatches

aws.agentcore.policy.target_resource.id

AgentCore Gateway resource identifier the request applies to

aws.agentcore.gateway.policy.arn

Policy Engine Amazon Resource Name (ARN) configured on the AgentCore Gateway

aws.agentcore.gateway.policy.mode

Policy Engine enforcement mode configured on the AgentCore Gateway, valid values are LOG_ONLY and ENFORCE

aws.agentcore.policy.guardrails.<category>.scores

Guardrail findings as Policy identifier, filter, and score tuples. Valid categories are contentFilter, promptAttack, and sensitiveInformation

aws.agentcore.policy.types

Policy identifier and policy type tuples for determining policies. Valid policy types are Cedar and Guardrail

aws.agentcore.policy.effects

Policy identifier and effect tuples. Valid effects are PERMIT, FORBID, and SuppressOutput

aws.agentcore.policy.guardrails.latency_ms

Time spent evaluating guardrails, in milliseconds

aws.agentcore.policy.log_only_matched_policies

List of LOG_ONLY Policy identifiers that matched the request

aws.agentcore.policy.log_only_decision_flipping_policies

List of LOG_ONLY Policy identifiers that change the authorization decision when promoted to ACTIVE

aws.agentcore.policy.log_only_mismatched_policies

List of LOG_ONLY Policy identifiers that failed due to missing attributes or type mismatches

aws.agentcore.policy.log_only_eval_incomplete

The string true when LOG_ONLY policy evaluation was incomplete. The attribute is omitted otherwise

aws.agentcore.policy.log_only_matched_policies.guardrails.<category>.scores

Guardrail findings for matched LOG_ONLY policies as Policy identifier, filter, and score tuples

aws.agentcore.policy.log_only_decision_flipping_policies.guardrails.<category>.scores

Guardrail findings for decision-flipping LOG_ONLY policies as Policy identifier, filter, and score tuples

aws.agentcore.policy.temporal.latency_ms

Time spent evaluating temporal policies, in milliseconds

aws.agentcore.policy.temporal.evaluation_invoked

Whether temporal policy evaluation ran for the request. This does not indicate that a temporal policy matched or determined the decision

aws.agentcore.policy.temporal.event_timestamp_ns

Exact event timestamp used by the temporal evaluator to order the request event, in nanoseconds

PartiallyAuthorizeActions

aws.agentcore.policy.allowed_tools

List of tool names that evaluated to an ALLOW decision

aws.agentcore.policy.denied_tools

List of tool names that evaluated to a DENY decision

aws.agentcore.policy.target_resource.id

AgentCore Gateway resource identifier the request applies to

aws.agentcore.gateway.policy.arn

Policy Engine Amazon Resource Name (ARN) configured on the AgentCore Gateway

aws.agentcore.gateway.policy.mode

Policy Engine enforcement mode configured on the AgentCore Gateway, valid values are LOG_ONLY and ENFORCE