Interface HarnessConfigProps

Tool allowlist controlling which tools (including the built-in shell / file_operations) the agent may select during invocation. Supports the AgentCore allowedTools patterns (*, plain names, @builtin, @server/tool, globs). Omit to allow all tools.

Note: allowedTools scopes LLM tool selection during InvokeHarness only - it does not gate the separate InvokeAgentRuntimeCommand API (which executes commands directly, without the LLM). To prevent direct command execution, do not grant bedrock-agentcore:InvokeAgentRuntimeCommand.

Use cases: restricting the default shell/file_operations tools, reducing tool-definition token overhead

AWS: AllowedTools

Validation: Optional; String[]; 1-64 entries

Inbound authorization configuration. Provide customJwt for JWT/OIDC inbound auth, or omit it to use AWS IAM (SigV4) - the Harness's no-configuration fallback.

Use cases: inbound access control

AWS: AuthorizerConfiguration

Validation: Optional; valid customJwt when present

Bring-your-own container image for the harness's underlying runtime environment (pre-built ECR image URI). Omit to use the AWS-managed harness container.

Use cases: custom runtime image

AWS: EnvironmentArtifact.ContainerConfiguration

Validation: Optional; HarnessContainerProperty

CloudWatch Data Protection configuration for the Harness's service-created log groups. PII masking and CMK encryption are always-on and cannot be disabled; this only tightens the posture by adding identifiers on top of the built-in floor.

Use cases: extending PII masking with additional identifiers

AWS: CloudWatch Logs Data Protection Policy

Validation: Optional; DataProtectionProperty; additive only

Named, versioned invocation endpoint for the harness. Omit to invoke the harness's default (latest) version directly.

Use cases: pinning callers to a specific harness version, blue/green endpoint management

AWS: AWS::BedrockAgentCore::HarnessEndpoint

Validation: Optional; HarnessEndpointProperty

Key-value environment variables passed to the harness runtime environment.

Use cases: runtime configuration, environment customization

AWS: EnvironmentVariables

Validation: Optional; Record<string, string>

Guardrail association for content filtering on the Harness's model calls. Rendered via a CDK property-override escape hatch (the pinned CDK L1 lags the CloudFormation spec for this field); see the class-level documentation.

Use cases: responsible-AI content filtering, safety controls

AWS: Model.BedrockModelConfig.AdditionalParams.guardrailConfig (escape hatch)

Validation: Optional; HarnessGuardrailAssociation

Idle-session and runtime lifecycle settings for the underlying AgentCore Runtime environment.

Use cases: cost control, session cleanup

AWS: Environment.AgentCoreRuntimeEnvironment.LifecycleConfiguration

Validation: Optional; HarnessLifecycleProperty

CloudWatch Logs retention period for the Harness's service-created log groups, in days. Accepts any CloudWatch Logs RetentionDays value; 9999 (RetentionDays.INFINITE) means never-expire and can be set explicitly to lock indefinite retention into config. Omitting the field is equivalent to 9999 - no retention policy is applied, leaving the log groups at CloudWatch's never-expire default (logs are kept, and billed, forever) unless a finite value is set.

Use cases: log retention policy, cost management

AWS: CloudWatch Logs log group retention

Validation: Optional; Number; must be a valid RetentionDays value (9999 for never-expire) - validated at synth

Maximum number of iterations the agent loop can execute per invocation.

Use cases: bounding tool-call loops, cost control

AWS: MaxIterations

Validation: Optional; Integer >= 1

Global maximum tokens the harness may generate across the whole agent-loop invocation (distinct from modelConfig.maxTokens, which bounds a single model call). A hard cost cap on total generation per invocation.

Use cases: cost control, bounding total generation per invocation

AWS: MaxTokens

Validation: Optional; Number; >= 1

Model sampling configuration (temperature, top-p, max tokens). Mirrors the CFN HarnessBedrockModelConfig sampling fields.

Use cases: tuning response determinism, diversity, and length limits

AWS: Model.BedrockModelConfig

Validation: Optional; HarnessModelConfigProperty

Foundation model identifier for the agent loop. Accepts an on-demand model id, a cross-region (system) inference profile id, or a full foundation-model / system inference-profile ARN (resolved via resolveModelArn). Application inference profile ARNs (application-inference-profile/...) are rejected at synth: their underlying foundation model is not derivable, so the paired invoke grant an inference profile requires cannot be scoped.

Use cases: selecting the reasoning model for the agent loop

AWS: Model.BedrockModelConfig.ModelId

Validation: Required; String; not an application-inference-profile ARN

VPC network configuration for the harness's runtime sessions, placing them behind your own security groups and subnets for private access to internal resources. Required: MDAA enforces VPC network isolation for the harness (NetworkMode: VPC), mirroring the AgentCore Runtime construct - there is no public-network option.

Use cases: private access to internal resources, network isolation

AWS: Environment.AgentCoreRuntimeEnvironment.NetworkConfiguration (NetworkMode: VPC)

Validation: Required; HarnessNetworkProperty; 1-16 security groups and subnets

Existing IAM role reference for the Harness execution role. If omitted, MDAA creates a role trusting bedrock-agentcore.amazonaws.com, scoped to this account/harness.

Use cases: role reuse, centralized permission management

AWS: ExecutionRoleArn

Validation: Optional; MdaaRoleRef

Skills injected into the agent's context: filesystem paths to skill definitions baked into the runtime image. Only the path skill source is exposed (the CDK L1 types only path); git / S3 / awsSkills sources documented in the CloudFormation HarnessSkill schema are not yet supported.

Use cases: injecting curated instruction/script bundles into the agent

AWS: Skills

Validation: Optional; HarnessSkillProperty[]; each path non-empty

System prompt defining the agent's behavior and instructions.

Use cases: agent persona, task instructions, behavioral constraints

AWS: SystemPrompt (one text block)

Validation: Required; String

Maximum duration in seconds for the agent loop execution per invocation.

Use cases: bounding total invocation latency

AWS: TimeoutSeconds

Validation: Optional; Integer >= 1

Tools available to the agent loop, keyed by tool name. The key becomes HarnessTool.Name and is what allowedTools entries refer to.

Use cases: giving the agent callable tools (client-executed or gateway-fronted)

AWS: Tools

Validation: Optional; NamedHarnessToolProps (map of tool name to config)

Context-truncation configuration controlling how the agent loop trims conversation context when it exceeds the model's context window. Omit to accept the service default.

Use cases: bounding long agent loops within the model context window

AWS: Truncation

Validation: Optional; HarnessTruncationProperty; tuning fields must match the selected strategy