At the AWS re:Invent 2025 conference in Las Vegas, Amazon Web Services announced transformative upgrades to its AI agent platform, Amazon Bedrock AgentCore, addressing critical enterprise concerns about safety, memory, and evaluation. The cloud giant revealed three major enhancements: Policy in AgentCore for natural language boundary controls, AgentCore Memory for persistent user preference tracking, and AgentCore Evaluations with 13 pre-built monitoring systems.
David Richardson, vice president of AgentCore, explained that the Policy feature allows developers to set boundaries using natural language instructions integrated with AgentCore Gateway, automatically checking and stopping actions that violate controls. This addresses enterprise fears about autonomous AI systems, with applications in access controls, third-party integration limits, and financial transaction boundaries.
The Memory capability enables AI agents to maintain detailed logs of user preferences and historical interactions, moving beyond simple task execution to become adaptive digital assistants. For example, it can remember travel preferences or customer support history to inform future decisions.
The Evaluation tools provide a suite of 13 pre-built systems to monitor correctness, safety, tool selection accuracy, and performance benchmarks. Richardson emphasized that this addresses deployment fears by offering standardized metrics without the tedium of building such systems from scratch.
These announcements signal a maturing AI agent market focused on enterprise requirements, with AWS positioning itself as a leader by enhancing control, personalization, and transparency. The event also featured keynotes from AWS CEO Matt Garman and other executives, streamed via platforms like AWS OnAir and Fortnite, highlighting broader cloud infrastructure innovations expected to impact various industries, including cryptocurrency.
