According to TechCrunch, AWS announced several new features for its Amazon Bedrock AgentCore platform during its annual AWS re:Invent conference. The company introduced Policy, a tool that lets developers set boundaries for AI agents using natural language, like restricting access to Salesforce or capping automatic refunds at $100. They also launched AgentCore Evaluations, a suite of 13 pre-built systems to monitor agent correctness and safety. Furthermore, AWS is building a memory capability called AgentCore Memory, which lets agents log user information over time to inform future decisions. David Richardson, vice president of AgentCore, stated these tools address the “biggest fears” about deploying agents and provide a head start on tedious evaluation work.
Agent controls are the real story
Here’s the thing: everyone’s building AI agents right now. The real differentiator isn’t just making them, it’s making them safe and manageable for big companies. That Policy feature is the sleeper hit. Telling an agent in plain English, “You can handle refunds under $100, but anything bigger needs a human,” is basically giving a non-technical business manager a seat at the table. It moves the conversation from pure coding to governance. And that’s huge for adoption. Without these guardrails, agents are just fancy, unpredictable chatbots that no enterprise risk officer would ever sign off on.
Why memory and evaluation matter
So the other two pillars—Memory and Evaluations—are about making agents actually useful and trustworthy over time. Memory turns a single interaction into a relationship. An agent that remembers your flight time or hotel preference isn’t just answering a question; it’s providing a service. But that’s also creepy if not handled right, which is why Policy has to come first. The Evaluations suite is AWS’s attempt to solve a massive pain point. Richardson called it “tedious to build,” and he’s not wrong. Providing 13 pre-built checks for safety and correctness gives teams a starting point they can customize. It’s a way to shortcut the fear of the unknown. Will your agent go off the rails? Well, here are 13 ways to check.
Is this just another trend?
Now, Richardson addressed the elephant in the room: are AI agents just a passing fad? His argument is that the core pattern—combining a model’s reasoning with the ability to take real-world actions via tools—is sustainable. I think he’s probably right. The hype will fade, but the functional need won’t. The tools and methods will change, which is why AWS is building a platform. They’re not selling a single agent; they’re selling the plumbing, guardrails, and monitoring tools for whatever agent architecture comes next. It’s a bet on the category, not a specific implementation. For enterprises dipping their toes in, that platform approach is less risky than betting on a single, monolithic agent product. You can learn more about the broader re:Invent event here.
