AWS unveils AI agent upgrades with security, memory, and testing tools at re:Invent
AWS has unveiled major updates to its Amazon Bedrock AgentCore platform at the re:Invent conference in Las Vegas. The new features—Policy, Evals, and episodic memory—aim to tackle key challenges in adopting AI agents for business use.
The changes will help companies deploy agents more securely while improving personalisation and performance tracking.
Amazon Bedrock AgentCore is a platform designed for building, managing, and scaling AI agents on AWS. At the conference, the company introduced three core additions to strengthen its capabilities.
Policy acts as a security and enforcement layer, setting guardrails for AI interactions. It operates outside the agent’s decision-making loop and uses natural language rules to control actions. This feature complements AWS’s existing security measures, including session-level isolation, to create a layered defence system.
Evals provides a way to test and refine AI agents through custom evaluations. Developers can choose from 13 prebuilt tests or design their own using an LLM as a judge. The tool supports both agent development and long-term operational monitoring.
Episodic memory extends the platform’s existing memory functions by storing user-specific preferences. This allows agents to recall individual interactions, improving personalisation over time.
David Richardson, VP of AgentCore at AWS, highlighted the Visa-AWS partnership as an example of these tools in action. The collaboration aims to streamline fragmented commerce systems by enabling secure, autonomous communication between agents.
The updates give businesses more control over AI agent behaviour while enhancing security and customisation. Policy, Evals, and episodic memory are now available within Amazon Bedrock AgentCore. Companies can use them to build, test, and deploy agents with greater confidence.
