Authoring Dogwood policies from natural language in Amazon Bedrock AgentCore

AI agents can automate complex workflows but may take actions that do not align with organizational policies or regulations if not properly controlled. Amazon Bedrock AgentCore addresses this by introducing a Policy feature that enforces controls across agents, now including time-based restrictions. These policies are expressed in the open-source language Dogwood and are applied in real time by the Dogwood monitor integrated into the AgentCore Gateway.

The recent expansion of Policy Authoring, an AI-driven tool, allows users to import natural language policy documents directly into Amazon Bedrock AgentCore. This facilitates the creation of policies that include temporal constraints, content filtering via Amazon Bedrock Guardrails, and restrictions on tool input parameters. The process involves decomposing policy rules, translating them into Dogwood formulas, and validating them against the environment schema.

Examples demonstrate how natural language policies, such as limiting refunds to business hours or requiring prior verification for transfers, are converted into precise Dogwood formulas. These formulas enforce rules based on time windows, event history, and content analysis, providing predictable and enforceable controls.

Best practices for policy construction emphasize clarity, explicitness, and specificity, such as stating whether the policy refers to attempts or outcomes, defining time windows explicitly, and naming the relevant subjects and thresholds. The translation pipeline ensures that only expressible rules are converted, with inexpressible rules flagged for alternative handling.

Overall, the new capabilities in Amazon Bedrock AgentCore streamline the process of creating, validating, and deploying formal policies from natural language, supporting organizations in maintaining safe and compliant AI agent behaviors.