AWS is expanding the capabilities of its Amazon Bedrock AgentCore platform by integrating automated reasoning, a math-based verification method, to enhance safety and control in AI agents. Announced at AWS’s re:Invent conference, the platform now includes new features: policy enforcement, evaluations, and episodic memory. These updates aim to provide enterprises with stronger controls over AI behavior and performance. The policy feature serves as a safeguard against agents breaking enterprise rules, operating independently from the agents’ core logic to prevent manipulation such as prompt injection attacks. Episodic memory allows agents to recall relevant but infrequently used information, improving the continuity and personalization of interactions. The evaluation tools offer customizable monitoring to maintain agent quality.