Imagine unleashing AI agents that can think, act, and evolve on their own—but only within the safe confines you've set, learning from every triumph and blunder to deliver smarter results over time. That's the thrilling breakthrough Amazon has just rolled out, and it's poised to revolutionize how we build and deploy artificial intelligence in the real world. Buckle up, because this isn't just about smarter tech; it's about ensuring AI stays trustworthy, adaptable, and downright useful without spiraling into chaos. But here's where it gets controversial: how much freedom should we grant these digital minds before they start making decisions that could reshape our businesses—or even our society? Let's dive in and unpack this game-changing announcement, step by step, so even newcomers to AI can follow along easily.
Amazon has unveiled cutting-edge AI Agents through its Amazon Bedrock platform, designed to operate responsibly, monitor their own effectiveness, and continually enhance their capabilities. For the latest updates, check out Amazon's News section (https://www.aboutamazon.com/amazon-news-today) and dive deeper into AWS innovations (https://www.aboutamazon.com/news/aws).
New Amazon Bedrock AgentCore features are fueling the evolution of agentic AI, enabling users to create, launch, and expand AI agents that are ready for real-world use with unparalleled reliability.
Authored by the Amazon Team
Time to read: About 5 minutes
Key highlights:
- The Policy feature in Amazon Bedrock AgentCore proactively prevents unauthorized actions by applying real-time, predictable safeguards that function independently of the agent's core programming.
- AgentCore Evaluations empowers creators to regularly assess an agent's performance quality through its real-world actions.
- AgentCore Memory adds episodic learning, allowing agents to draw from past experiences to refine their choices.
- Companies across various scales and compliance needs—including Amazon Devices Operations & Supply Chain, Archera.ai, Cohere Health, Cox Automotive, Druva, Heroku, Natera, NTT Data, MongoDB, PGA TOUR, Pulumi, Thomson Reuters, Workday, Snorkel.ai, Swisscom, and S&P Global Market Intelligence—rely on AgentCore to swiftly bring their AI agents to production status.
Today marks a significant milestone as Amazon introduces fresh advancements in Amazon Bedrock AgentCore (https://aws.amazon.com/bedrock/agentcore/), the premier toolset for constructing and implementing AI agents securely and at scale. The Policy component lets teams define strict limits on agents' interactions with tools, while AgentCore Evaluations provides insights into how these agents will fare in practical scenarios. Plus, AWS has upgraded its memory systems to let agents absorb lessons from prior engagements, delivering more personalized responses to users over time. Explore more on AWS (https://www.aboutamazon.com/amazon-aws-news).
Craft enterprise-grade AI agents that recognize their strengths and boundaries
AI agents' knack for independent reasoning and execution is undeniably potent, yet businesses need solid safeguards to avert unauthorized data exposure, improper engagements, or operational errors that might disrupt daily workflows. Despite meticulous guidance, these agents sometimes err in reality, leading to potentially grave repercussions.
Enter Policy in Amazon Bedrock AgentCore, a tool that empowers organizations to establish explicit guidelines for agent behavior. Teams can articulate rules in everyday language, specifying accessible tools, permissible actions, and the circumstances under which they apply. These might include APIs, Lambda functions, MCP servers, or widely-used services from partners like Salesforce and Slack. To keep things speedy, Policy integrates with AgentCore Gateway, verifying actions against rules in mere milliseconds. This keeps agents autonomous yet compliant. The user-friendly, language-based policy creation simplifies the process, letting folks describe guidelines conversationally rather than coding them formally. Picture a straightforward rule such as 'Deny all refunds exceeding $1,000 per customer.' This embodies Amazon's mantra of 'trust, but verify,' enabling agents to function independently under watchful eyes.
And this is the part most people miss: in a world where AI might one day handle critical decisions, is this level of oversight enough to prevent biased or unethical choices? Could overly strict policies stifle innovation, or do they protect us from AI gone rogue? Druva, a top data protection expert, shares their perspective. 'Customers often spend countless hours sifting through logs from multiple systems when backup failures occur,' explains David Gildea, Vice President of Product AI at Druva (https://www.aboutamazon.com/artificial-intelligence-ai-news). 'With our AI agents (https://www.aboutamazon.com/news/aws/aws-summit-agentic-ai-innovations-2025), they now receive immediate diagnostics and guided fixes for data restoration. We're thrilled to adopt Policy in AgentCore, as it'll let us restrict agent access to sensitive resources like backup tools, security records, and monitoring panels. With solid policies, our engineers can experiment boldly, assured that compliance is upheld. This lets us grow our agent ecosystem while meeting the rigorous security demands of our enterprise clientele.'
Achieve full transparency into AI agent actions and outcomes
Assessing AI agent performance isn't like checking traditional software; it demands intricate data analysis, subjective judgments, and ongoing oversight, a hurdle that intensifies with every tweak or model shift.
AgentCore Evaluations streamlines this by offering 13 ready-made assessment tools for essential metrics like accuracy, usefulness, tool choice precision, security, success rates, and contextual fit. Creators can also design custom checks using their favorite large language models and prompts. What once took teams months of data expertise now happens seamlessly, with live interactions sampled for metrics such as correctness, helpfulness, and safety. Teams can configure notifications for quality tracking during tests and live operations. For instance, if a support agent's satisfaction ratings plummet by 10% in eight hours, alerts prompt quick fixes to protect user experience.
But here's where it gets controversial: while these evaluations promise reliability, could they mask underlying biases in AI decision-making? Are we truly measuring 'success' if it favors certain demographics or ignores ethical nuances? Natera, a pioneer in genetic testing and healthcare diagnostics, weighs in. 'At Natera, we're revolutionizing cancer care with AI agents,' says Mirko Buholzer, Software Engineering Lead at Natera. 'Our squads are deeply invested in ensuring uniform quality and effectiveness across our AI systems, all while adhering to healthcare regulations. AgentCore Evaluations will be instrumental, offering ongoing performance monitoring via key indicators like precision, assistance value, and user contentment. This live intelligence should help us spot and resolve issues early. With these tools, we're poised to launch dependable agents that uphold our benchmarks and facilitate life-changing care at a grand scale.'
Develop agents that grow wiser with each encounter
Today's AI agents often fall short on memory, stuck in short-lived conversation windows that erase lessons after each exchange, hindering growth from production successes or setbacks.
AgentCore Memory fills this gap, letting agents form a lasting user profile over time. Now generally available, its episodic feature enables agents to glean from historical events and inform future choices. Episodes record details like context, logic, steps taken, and results, with agents analyzing patterns for better judgments. Facing similar tasks, they retrieve pertinent past data swiftly, cutting down on processing and custom setups. Consider an agent scheduling airport transport 45 minutes before a solo flight. Months later, heading to the same spot with family, it instinctively books pickup two hours early, recalling familial travel hurdles. This data-driven learning fosters steadier choices over rigid rules.
S&P Global Market Intelligence, offering data and tech solutions to investors and firms, shares their story. 'We built Astra, our versatile agent workflow system, but coordinating intricate multi-agent tasks across our spread-out team was tough,' notes Astier Helen, Head of Technology at MI Enterprise Technology and Sustainability, S&P Global. 'As specialized agents multiplied, tracking state and context became a nightmare, underscoring the need for a shared memory system. Amazon Bedrock AgentCore Memory delivered with integrated, centralized checkpoints for our multi-agent setups. The new episodic memory lets our agents build on past analyses for sharper insights. Before, agent deployment dragged on for weeks; now, with AgentCore, we whip up and launch an agent or MCP server in minutes.'
These latest developments provide tailored infrastructure for agent creation, freeing you to innovate instead of reinventing AI basics.
For in-depth info on the new Amazon Bedrock AgentCore features, head to:
Stay updated on AWS re:Invent (https://www.aboutamazon.com/aws-reinvent-news-updates), covering agentic and generative AI, new products, and more.
Additional Amazon Updates
1 / 1
What do you think? Is Amazon striking the perfect balance between AI autonomy and safety, or are we on the brink of over-regulating technology that could otherwise accelerate progress? Could these memory features lead to AI that 'remembers' our preferences too well, raising privacy concerns? Share your thoughts in the comments—do you agree, disagree, or have a counterpoint we've missed? Let's keep the conversation going!
)
