AWS Bedrock Adds Temporal Policies as AI Agent Control Moves Beyond the Model

AWS this week expanded Amazon Bedrock AgentCore with temporal policies and an open-source policy language called Dogwood, enabling AI agent gateways to evaluate sequences of actions rather than individual tool calls in isolation. The update addresses scenarios where each single action appears compliant but the cumulative pattern violates organizational intent, such as an agent making multiple near-limit purchases. The architectural shift places runtime control outside the model itself, meaning agents no longer rely on faithfully recalling constraints from their prompts. A parallel trend is emerging across the ecosystem, with platforms like Tenable's CyberAgents Exchange decomposing agents into discrete, reusable components including skills, plans, and policies. However, analysts note that runtime policy enforcement still leaves open a deeper question: determining whether the underlying organizational judgment or evidence actually warrants an approval, a problem distinct from simple tool authorization.