AWS uses automatic reasoningthat uses math-based verification to create up-to-date opportunities on your Amazon Bedrock AgentCore platform as the company dives deeper into the agentic AI ecosystem.
Announced at the annual re:Invent conference in Las Vegas, AWS adds three up-to-date capabilities to AgentCore: “policy,” “ratings,” and “episodic storage.” The up-to-date features are intended to give enterprises greater control over agent behavior and performance.
AWS also revealed what it calls a “new class of agents” or “border agents” that are autonomous, scalable, and independent.
Swami Sivasubramanian, AWS vice president of Agentic AI, told VentureBeat that many of AWS’s up-to-date features represent a shift in who becomes a builder.
“We’re actually on the cusp of a major tectonic shift around AI, but agentic AI is really starting to transform what is the art of the possible and will make it one of the truly transformative technologies,” Sivasubramanian said.
Political agents
Up-to-date pageThe olicy feature helps enterprises reinforce guidance even when the agent has already justified its response.
AWS Vice President of AgentCore David Richardson told VentureBeat that the policy tool sits between the agent and the tools it calls, rather than being baked into the agent as is often the case with tuning. The idea is to prevent the agent from violating company policy and redirect him to re-evaluate his reasoning.
Richardson gave the example of a customer service agent: The company would write a policy stating that the agent could issue a refund of up to $100, but for an amount higher than that, the agent would have to refer the customer to a human. He noted that it is straightforward to subvert an agent’s reasoning loop, for example by instant data injection or poisoning, leading agents to ignore guardrails.
“There are always instant injection attacks where people try to undermine the agent’s reasoning to get it to do things it shouldn’t do,” Richardson said. “That’s why we’ve implemented this policy outside of the agent, and it works using the automated inference capabilities we’ve built over the years to help customers determine their options.”
AWS presented Automatic reasoning checking on Bedrock at last year’s re: Invent. These utilize neurosymbolic AI or math-based verification to prove correctness. The tool applies mathematical evidence to models to confirm that hallucinations have not occurred. AWS relies heavily on neurosymbolic AI and automated reasoning, pushing for enterprise-grade security in a way that is different from other AI model providers.
Episodic memories and evaluations
The other two up-to-date updates to AgentCore, “ratings” and “episodic memory,” also give enterprises greater visibility into agent performance and provide agents with episodic memory.
ANDIn the case of AgentCore’s memory enhancement, episodic memory refers to knowledge that agents utilize only occasionally, as opposed to long-term preferences that they must constantly refer to. The limitations of the context window make things complex for some agents, sometimes causing them to forget information or conversations they haven’t used for a while.
“The idea is to help capture information that the user would really want the agent to remember when they come back,” Richardson said. “For example: ‘What is their favorite seat on the plane when traveling as a family?’ Or “what is the price range they are looking for?”
Episodic memory differs from previously provided AgentCore memory because instead of relying on short- and long-term memory maintenance, agents built on top of AgentCore can recall specific information based on triggers. This may eliminate the need for custom instructions.
With AgentCore evaluations, organizations can use 13 ready-made evaluators or write their own. Developers can set alerts to warn them if agents start to fail in quality monitoring.
Border agents
But perhaps AWS’s strongest push into enterprise agent AI is the release of border agents, which are fully automated and independent agents that the company says can act as team members with little direction.
The concept is similar, if not identical, to the concepts of more asynchronous competitor agents Google AND OpenAI. However, AWS appears to provide more than just autonomous coding agents.
Sivasubramanian called them a “up-to-date class” of agents, “not just a step change in what you can do today; they’re moving from helping with individual tasks to sophisticated projects.”
The first is Kiro, an autonomous coding agent that has been available in public preview since July. At the time, Kiro was touted as an alternative to vibration encoding platforms like OpenAI Code Or Windsurfing. Like countless Codex and Google asynchronous coding agents, including JulesKiro can code, perform reviews, fix bugs on his own, and specify tasks to perform.
Meanwhile, the AWS Security Agent embeds deep security knowledge into applications from the very beginning. The company said in a press release that users “define security standards once, and the AWS Security Agent automatically checks them across applications during review, helping teams address risks relevant to their business rather than generic checklists.”
The AWS DevOps agent will support developers, especially those on duty, proactively find system failures and errors. Can respond to incidents using their knowledge of the application or service. It also confirms links between the application and the tools it uses, such as Amazon CloudWatch, Datadog, and Splunk, to trace the root cause of the problem.
Enterprises are interested in deploying agents and ultimately incorporating more autonomous agents into their workflows. And while companies like AWS continue to provide these agents with security and control, organizations are slowly figuring out how to connect them all.