Amazon Web Services (AWS) is strengthening its AI agent platform, Amazon Bedrock AgentCore, with a series of updates designed to make building, managing, and monitoring AI agents more robust for enterprise users. The new features were announced at AWS’s annual re:Invent conference and aim to address key concerns around AI agent safety, memory, and operational boundaries.
One major addition is AgentCore Policy, a tool that allows developers to set natural language-based boundaries for AI agent actions. Integrated with AgentCore Gateway, which connects AI agents to external tools, Policy automatically enforces rules by halting actions that violate set constraints. For instance, developers can restrict agents to accessing only certain internal data or third-party apps like Salesforce or Slack. Businesses can also configure agents to perform specific tasks autonomously, such as issuing refunds up to a set amount, while requiring human oversight for larger transactions. This approach is intended to mitigate risks and ensure agents operate within safe and predictable parameters.
Another significant update is AgentCore Evaluations, which introduces 13 pre-built evaluation systems that monitor key performance metrics, including correctness, safety, and tool selection accuracy. These tools give developers a foundation for creating custom evaluation systems, saving significant development time while addressing concerns around agent reliability. “This feature helps tackle one of the biggest fears associated with deploying AI agents,” said David Richardson, vice president of AgentCore.
AWS is also enhancing the platform with AgentCore Memory, enabling AI agents to track user preferences and interactions over time. This memory can include details such as travel preferences, meeting schedules, or workflow patterns, allowing agents to provide more personalized and contextually informed responses.
Richardson emphasized that these updates are designed to support enterprises in leveraging AI agents sustainably, even as the industry evolves. “Being able to combine reasoning capabilities with real-world tool execution is a model that will remain relevant,” he said. “The specifics of the tools will change, but the underlying pattern is durable.”
With these enhancements, AWS positions AgentCore as a leading enterprise solution for developing AI agents capable of safely interacting with systems, retaining user context, and being continuously evaluated for performance and compliance.
