AWS adds natural language policy authoring to Bedrock AgentCore

AWS built Policy in Amazon Bedrock AgentCore to address a core challenge: AI agents operating without proper controls can take actions that violate organizational policies or regulatory requirements. The new Policy Authoring tool directly addresses this by bridging the gap between how compliance teams think (natural language rules and procedures) and how formal governance systems work (structured logic languages). This is particularly important for regulated industries where documented controls are mandatory but the technical implementation often falls to specialized engineers.

The tool's design reflects a pragmatic understanding of how policies already exist in organizations: compliance teams maintain written documents for human staff, and those same documents contain the logic that should govern automated systems. By accepting these natural language documents as input and translating them to Dogwood, Policy Authoring eliminates the manual transcription step that would otherwise require either compliance teams to learn a formal language or engineers to interpret prose into logic. The examples in the blog post—refunds during business hours and capped at $2,500, transfers requiring prior identity verification within 15 minutes, cumulative transfer caps of $50,000 in 12 hours—show common financial controls that map directly to tool arguments, temporal sequences, and aggregate constraints. Each demonstrates that the tool handles not just simple restrictions but nuanced rules that refer to prior actions and cumulative effects across a session.