Amazon Web Services (AWS) is enhancing its AI agent platform, Amazon Bedrock AgentCore, to make it easier for enterprises to build and monitor AI agents.
AWS announced several new AgentCore features on Tuesday during the company’s annual AWS re:Invent conference. The company announced new tools for managing AI agent boundaries, agent memory capabilities, and agent evaluation capabilities.
One of the upgrades is the introduction of policies to AgentCore. This feature allows users to use natural language to set the boundaries of their interactions with agents. These boundaries are integrated with AgentCore Gateway, which connects AI agents to external tools and automatically checks each agent’s actions and stops those that violate written controls.
Policies allow developers to set access controls to specific internal data or third-party applications such as Salesforce or Slack. Those boundaries could include telling AI agents that they can automatically refund up to $100, but beyond that they need to involve a human, David Richardson, vice president at AgentCore, told TechCrunch.
The company also announced AgentCore Evaluations, a suite of 13 pre-built evaluation systems for AI agents that monitor factors such as accuracy, safety, and precision of tool selection. This also gives developers a head start in building their own evaluation capabilities.
“This is going to really help address the biggest fear people have when deploying agents,” Richardson said of the new rating feature. “(It’s) something that a lot of people want to have, but it’s a hassle to make.”
AWS also announced that it is incorporating memory capabilities into its agent platform, AgentCore Memory. This feature allows agents to log information about users over time, such as flight times and hotel preferences, and use that information to make future decisions.
tech crunch event
san francisco
|
October 13-15, 2026
“We continue to iterate on these three things at different layers of AgentCore,” says Richardson. “We use policies to communicate with existing systems, use (AgentCore memory) to make (agents) more powerful, and allow development teams to iterate on agents.”
Agents are currently the mainstay of the AI industry, but some believe this technology won’t last long. But Richardson believes the tools AgentCore is developing can withstand a rapidly changing market, even if trends change, and he hopes they will.
“Being able to leverage the inference capabilities of these models, coupled with being able to do real-world things through the tools, feels like a sustainable pattern,” Richardson said. “The way that pattern works is definitely going to change. I think we’re ready for that.”
Check out all of our coverage of TechCrunch’s annual enterprise technology event here.
