Amazon Web Services (AWS) has expanded its artificial intelligence (AI) agent platform, Amazon Bedrock AgentCore, with a range of new tools that make managing, building, and assessing AI agents simpler for enterprises.
The updates were revealed on Tuesday at AWS re:Invent, aiming to push the boundaries towards excellence.
The significant addition is Policy in AgentCore that allows developers to set natural-language boundaries for what an AI agent can and cannot do.
These rules incorporate with AgentCore Gateway to automatically check agent actions and inhibit anything that violates written controls.
A company could enable an assistant to approve refunds of nearly $100 but need a human agent to approve refunds up to $100 but require human approval for larger amounts.
AgentCore Evaluations
AWS has also launched AgentCore Evaluations, a comprehensive suite of 13 pre-built systems that assess agents on correctness, safety, tool-use accuracy, and more.
According to the company, it will assist teams deploy agents more confidently without building evaluation tools from scratch.
AgentCore Memory
Another update, AgentCore Memory, enables assistants to store information about users — such as travel details or personal preferences — and use it to make better decisions over time.
According to David Richardson, VP of AgentCore, these three upgrades strengthen the platform across boundaries, memory, and safety.
While some critics believe AI agents are a passing trend, Richardson argues the combination of strong reasoning models and tool-based actions is a “sustainable pattern,” even with a transition in industry.
