Amazon Bedrock AgentCore turns natural language into Dogwood governance policies

Amazon Bedrock AgentCore now includes Policy Authoring, an AI-driven tool that converts natural language policy documents into syntactically and semantically correct Dogwood formal specifications. Dogwood is an open source governance language that runs in the AgentCore Gateway monitor, enforcing restrictions on agent actions in real time. The new capabilities extend beyond basic input parameter restrictions to temporal and trajectory constraints, enabling policies for rate limiting, prerequisite and sequential ordering of tool calls, and cumulative effects. It also integrates Amazon Bedrock Guardrails to detect inappropriate content in free-form text.

Policy Authoring works best when you have clean, rule-focused prose—like a policies list or an operating procedure's rules section—rather than documents heavy on rationale. It acts as a translator, not a summarizer. The tool automatically pulls the agent's MCP tool manifest to build a schema, so generated policies reference the exact tool names, arguments, and return values the agent uses. For example, a retail bank's customer-servicing agent can enforce rules such as verifying identity before initiating transfers, requiring supervisor approval for charges, or capping refund amounts—all written in plain English and converted into Dogwood's default-deny, permit-override semantic model.