Amazon Web Services is stepping up its push into enterprise AI with a major expansion of its AgentCore platform, unveiled at the AWS re:Invent conference. The company introduced a fresh set of tools designed to make it easier for organisations to build, manage and monitor AI agents.
The upgrades focus on three areas: stricter policy controls, improved agent evaluation, and a new memory system meant to help AI agents deliver more personalised and reliable outcomes.A key highlight is the introduction of Policy in AgentCore, which lets developers set clear boundaries for what an AI agent can or cannot do using simple natural language instructions.
These rules integrate with the AgentCore Gateway to automatically check every action an AI agent takes and block any that violate set controls. AWS says this gives enterprises greater confidence, especially when agents must interface with sensitive data or tools like Salesforce and Slack, or when handling tasks that require human oversight, such as approving refunds above certain limits.
AWS also rolled out AgentCore Evaluations, a suite of 13 pre-built evaluation systems that monitor an agent’s accuracy, safety, decision-making and tool-usage patterns. The goal is to help developers quickly establish reliable guardrails without building everything from scratch.
According to David Richardson, vice president of AgentCore, the new evaluation tools address one of the biggest concerns businesses have about deploying AI agents: ensuring they behave consistently and safely in real-world environments. Rounding off the updates is AgentCore Memory, a new capability that allows AI agents to retain useful information about users such as travel preferences or frequently requested tasks and apply that knowledge in future interactions.
Richardson believes this combination of policy controls, memory and evaluation features positions AgentCore to remain relevant even as AI trends shift. He argues that blending strong reasoning capabilities with the ability to perform real-world tasks is a sustainable model for the future, regardless of how quickly the AI landscape evolves