Amazon Web Services (AWS) is significantly enhancing its Amazon Bedrock AgentCore platform, aiming to simplify the development and monitoring of AI agents for enterprises. The company unveiled several new AgentCore features at its annual AWS re:Invent conference on Tuesday, including advanced tools for managing AI agent boundaries, robust agent memory capabilities, and comprehensive agent evaluation systems.
A key new capability is Policy in AgentCore, which empowers users to define interaction boundaries for AI agents using natural language. This feature integrates seamlessly with AgentCore Gateway, the component that connects AI agents to external tools, to automatically scrutinize each agent’s actions and halt any that breach predefined controls. Developers can leverage Policy to establish access controls for sensitive internal data or third-party applications such as Salesforce or Slack.
David Richardson, Vice President of AgentCore, explained to TechCrunch that these boundaries could also dictate that an AI agent automatically processes refunds up to $100, but requires human intervention for larger amounts.
AWS also introduced AgentCore Evaluations, a comprehensive suite of 13 pre-built evaluation systems designed for AI agents. These systems monitor crucial factors such as correctness, safety, and the accuracy of tool selection, among others. This offering provides developers with a significant head start in building their own custom evaluation features.
“That one is really going to help address the biggest fears that people have [with] deploying agents,” Richardson said about the new evaluation capabilities. “[It’s] a thing that a lot of people want to have but is tedious to build.”
Furthermore, AWS is integrating a new AgentCore Memory capability into the platform. This feature enables AI agents to progressively build a log of user-specific information over time, such as flight times or hotel preferences, and then leverage this accumulated knowledge to inform future interactions and decisions.
Richardson summarized the strategic enhancements, noting, “Across these three things, we are continuing to iterate at the different layers at AgentCore: talking to existing systems with Policy, [making agents] more powerful with [AgentCore Memory], helping the development team iterate with an agent.”
Despite some industry skepticism regarding the long-term viability of AI agents, Richardson expressed confidence that AgentCore’s new tools are robust enough to adapt to the rapidly evolving AI market. He believes that “being able to take advantage of the reasoning capabilities of these models, which is coupled with being able to do real-world things through tools, feels like a sustainable pattern.” While acknowledging that “the way that pattern works will definitely change,” Richardson affirmed, “I think we feel ready for that.”
For more coverage of the annual enterprise tech event, you can follow along here.
