AWS announced a significant update to AgentCore, its platform for building and running AI agents, at the AWS Summit in New York today. For business operators, the practical question is simple: what can your AI agents actually do now that they could not do easily before?
The first problem AgentCore is solving is access. Most company knowledge sits in SharePoint folders, Confluence wikis, Google Drive, and internal PDFs. Getting an AI agent to read that material used to require months of custom engineering work before the agent could answer a single question about your own business. The new Managed Knowledge Base service handles all of that setup for you: connect your sources, and AWS manages everything else.
For information outside the company, AgentCore now includes a built-in web search tool that searches the live internet and brings results back inside your AWS security setup. No separate vendor to sign up for, no additional billing system. The search uses the same infrastructure that powers Amazon's Alexa and Quick products, with a proprietary Amazon knowledge graph on top for verified facts and real-time data like market prices.
For paid data, AgentCore Payments lets agents access licensed research, financial feeds, and premium APIs as part of their normal workflow. Content owners who want to charge agents for access can use a new AWS web protection tool to control and monetize that traffic. This is effectively the early infrastructure for agents buying things on your behalf.
The second, and arguably more important, problem is failure detection. The industry data here is sobering. According to McKinsey, nearly two-thirds of enterprises have experimented with AI agents, but fewer than 10% have scaled them successfully. Gartner projects that over 40% of agentic AI projects will be canceled by 2027. The core reason is not that the models are bad. It is that agents fail in ways that traditional monitoring cannot see.
An AI agent can confirm an order it never placed. It can skip an approval step while your dashboard shows a 99% success rate. It can give increasingly wrong answers for weeks before anyone notices. One documented case involved a customer service agent whose quality scores quietly dropped while standard monitoring showed no errors at all, and the problem only surfaced when customer satisfaction fell significantly over several weeks.
AgentCore's new monitoring analyzes hundreds of production sessions at once and surfaces recurring failure patterns, ranked by how many users they are affecting. The platform then suggests specific improvements to the agent's instructions and tool descriptions, and lets teams run a live split-test between the old and new version before making any changes permanent. This is the part that matters most: it turns agent improvement from guesswork into a structured process.
On security, the new Bedrock Guardrails integration runs checks for attempted manipulation of the agent, harmful content, and sensitive data exposure at a layer the agent itself cannot see or reason around. More security providers, including Check Point, Zscaler, and SentinelOne, are coming to that same layer soon.
For any organization already running AI agents or planning to, the monitoring and failure-detection features are the ones worth paying attention to first. The knowledge access tools reduce build time. The monitoring tools protect the business from the slow, invisible failures that have quietly ended most enterprise AI projects so far.