"One upgrade is the introduction of Policy in AgentCore. This feature allows users to set boundaries for agent interactions using natural language. These boundaries integrate with AgentCore Gateway, which connects AI agents with outside tools, to automatically check each agent's action and stop those that violate written controls. Policy allows developers to set access controls to certain internal data or third-party applications like Salesforce or Slack. These boundaries can also include telling an AI agent they can automatically issue refunds up to $100 but must bring a human in the loop for anything larger, David Richardson, vice president of AgentCore, told TechCrunch."
"The company also announced AgentCore Evaluations, which is a suite of 13 pre-built evaluation systems for AI agents that monitor factors including correctness, safety, and tool selection accuracy, among others. This also allows developers to have a head start in building their own evaluation features as well. "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.""
AWS expanded AgentCore with Policy controls, AgentCore Gateway integration, AgentCore Evaluations, and AgentCore Memory. Policy enables natural-language boundaries that automatically check and stop agent actions that violate written controls when integrated with the Gateway. Policy supports access controls to internal data and third-party apps and allows granular rules such as automated refunds under $100 while requiring human approval for larger amounts. AgentCore Evaluations delivers 13 prebuilt evaluation systems measuring correctness, safety, and tool-selection accuracy and provides a foundation for custom evaluations. AgentCore Memory records user-specific information over time to inform future agent decisions and personalization.
Read at TechCrunch
Unable to calculate read time
Collection
[
|
...
]