Authoring Dogwood policies from natural language in Amazon Bedrock AgentCore
aws ·
AWS announced expanded capabilities for Policy Authoring in Amazon Bedrock AgentCore, an AI-driven tool that translates natural language policy documents into Dogwood formal specifications. The generated policies can enforce temporal and trajectory constraints, invoke Amazon Bedrock Guardrails for content checks, and restrict tool input parameters. The post illustrates the feature with a retail bank customer-servicing agent whose tools include verify_identity, initiate_transfer, issue_refund, file_dispute, and request_approval. Dogwood is default-deny, so rules that grant capabilities become permits with conditions, while limits become forbids. Example policies cover business-hours refund restrictions, prerequisites, cumulative transfer caps, rate limits, and sensitive-information checks. The authors also provide best practices, such as clarifying whether a rule concerns attempted or completed actions and stating time windows explicitly.
aws