AWS Announces New Capabilities for Its AI Agent Builder

Amazon Web Services (AWS) is expanding its AI agent development platform, Amazon Bedrock AgentCore, introducing new features that simplify how enterprises build, manage, and monitor AI agents.

The updates were unveiled on Tuesday at the company’s annual AWS re: Invent conference. AWS introduced enhancements focused on managing agent boundaries, improving memory capabilities, and strengthening evaluation tools.

One notable addition is Policy in AgentCore, a feature that enables developers to define rules and constraints for agent behaviour using natural language. These rules work alongside AgentCore Gateway, which links AI agents to external tools and systems, to automatically review each action the agent takes and block any that violate established policies.

Policy settings allow organisations to restrict access to internal systems or third-party tools such as Salesforce or Slack. They can also set operational limits — for example, an AI agent may be allowed to issue refunds up to $100. Still, anything beyond that would require human approval, AWS vice president of AgentCore David Richardson explained to TechCrunch.

AWS also introduced AgentCore Evaluations, a suite of 13 prebuilt evaluation tools designed to monitor metrics such as correctness, safety, and tool selection accuracy. These evaluation modules provide developers with a foundation for building more customised oversight systems.

“That one is really going to help address the biggest fears that people have with deploying agents,” Richardson said. “[It’s] something many teams want but is tedious to create from scratch.”

The third major update is AgentCore Memory, which allows AI agents to build long-term memory by storing information about users — such as travel times, hotel preferences, and other behavioural data — and incorporating it into future decisions.

“Across these three things, we are continuing to iterate at the different layers of AgentCore,” Richardson said. “Talking to existing systems with Policy, making agents more powerful with AgentCore Memory, helping the development team iterate with an agent.”

While AI agents dominate industry discussions right now, some analysts believe the trend may fade. Richardson, however, is confident that AgentCore’s approach is built for longevity even as the market rapidly evolves.

“Being able to take advantage of the reasoning capabilities of these models, paired with the ability to take real-world actions through tools, feels like a sustainable pattern,” he said. “The way that pattern works will definitely change. I think we feel ready for that.”