AWS presents its new comprehensive AI strategy

Amazon Web Services (AWS) held its annual AWS re:Invent 2025 event and announced many innovations, ranging from devices and technological advances to collaborations with other companies.

This article will focus on three innovations: frontier agents for software development, the Nova 2 model family, and the new AgentCore capabilities in Amazon Bedrock.

These solutions are establishing a turning point in the way organizations implement and optimize artificial intelligence to improve their operational efficiency and application security. Below, we analyze each one of them.

The 3 AWS Frontier Agents

One of the most notable advancements in AWS’s artificial intelligence strategy is the introduction of frontier agents, a new class of artificial intelligence agents designed to work autonomously, at scale, and independently, optimizing the software development lifecycle. The autonomous Kiro, AWS Security Agent, and AWS DevOps Agent are aimed at enhancing development, security, and IT operations processes.

Kiro: a virtual developer that learns over time

The autonomous Kiro agent acts as a virtual developer that can manage tasks such as error classification and improving code coverage without requiring constant oversight. This agent learns continuously, retaining context across sessions and increasing team efficiency.

Developers may assign tasks directly from platforms such as Jira, GitHub, and Slack, and receive results without having to deal with friction such as coordinating between different repositories. Every code review, ticket, and architectural decision informs the agent’s understanding, making it increasingly useful over time.

AWS Security Agent

AWS Security Agent assists in building secure applications from the outset in AWS, multicloud, and hybrid environments. The agent proactively reviews design documents and analyzes pull requests to verify organizational security requirements and common vulnerabilities. The company itself defines its security standards once, and AWS Security Agent automatically validates them across all applications during review, helping teams address business-critical risks rather than relying on generic checklists.

In addition, it transforms penetration testing from a manual and slow process into an on-demand capability, adapting to the team’s development speed.

The DevOps Agent enhances operations and reliability

The AWS DevOps Agent is an active member of the operations team, working to proactively resolve incidents and enhance application reliability through continuous recommendations based on system behavior. It reduces incident response time and optimizes overall application performance.

As a use case, AWS highlights the case of the Commonwealth Bank of Australia, one of Australia’s leading providers of integrated financial services, with more than 17 million customers. “The bank’s Cloud Foundations group manages more than 1,700 AWS accounts and provides centralized cloud operations services to thousands of engineers. While developing the prototype for its next-generation internal cloud platform, the team replicated a complex network and identity management problem to test the AWS DevOps Agent. Such issues may take an experienced DevOps engineer hours to identify, and the agent found the root cause in less than 15 minutes,” the company explains.

Nova 2: the new era of reasoning and generative AI models

In parallel with the introduction of the agents, AWS has launched the Nova 2 model family, which includes several versions tailored to diverse business needs. We provide a brief analysis of each one below:

• Nova 2 Lite for everyday tasks: designed for fast and low-cost reasoning tasks, such as task automation, code generation, and customer support via chatbots. With an outstanding price-performance ratio, Nova 2 Lite outperforms models such as Claude Haiku 4.5 and GPT-5 Mini in various performance tests.
• Nova 2 Pro for complex challenges: the most advanced model in the series, ideal for tasks requiring greater precision, such as agentic coding, long-term planning, and the resolution of complex problems. This model stands out not only for its capacity to process text, images, and voice, but also for its ability to transfer knowledge to smaller models through knowledge distillation.
• Nova 2 Sonic: ideal for building conversational artificial intelligence applications that seamlessly integrate text and voice.
• Nova 2 Omni: enables the simultaneous processing of text, images, video, and voice, eliminating the need for several specialized models for complex tasks.
• Nova Forge: allows organizations to create custom artificial intelligence models, tailored to their specific needs through a process called open training. This service gives companies full control over the integration of proprietary data during model training, resulting in solutions more uniquely suited to their operations.
• Nova Act: facilitates the creation of reliable artificial intelligence agents that automate workflows in web environments. This service enables companies to rapidly prototype agents using natural language and then refine them in integrated development environments such as VS Code. The results may be implemented and scaled efficiently in production, thus allowing organizations to improve operational efficiency by automating repetitive and complex tasks.

Amazon Bedrock and its AgentCore capabilities

To complement this strategy, AWS has updated Amazon Bedrock with new capabilities in AgentCore, a platform for the creation, deployment, and scaling of artificial intelligence agents. AgentCore facilitates the creation of granular access policies and monitoring of agent behavior via continuous assessments.

These assessments provide full visibility into the behavior and results of artificial intelligence agents, exceeding the limitations of traditional metrics. The service simplifies the assessment processes through 13 prebuilt evaluators that analyze key aspects such as accuracy, usefulness, security, and objective achievement. Additionally, developers may create custom evaluators.

Unlike previous approaches that required months of effort, AgentCore provides continuous analytics of agents’ live interactions. Teams may configure alerts for proactive quality monitoring, which enables a rapid response to any drop in agent performance, thus ensuring an optimal user experience.

Control and security with Policy

The new Policy feature in AgentCore enables companies to establish specific boundaries using natural language for what agents may do, which tools and data they are permitted to access, and the conditions under which these may occur. These tools can include APIs, Lambda functions, MCP servers, or popular third-party services such as Salesforce and Slack.

To ensure agent responsiveness and speed, Policy is integrated with AgentCore Gateway to instantly compare agents’ actions with policies in milliseconds.

Episode memory to improve decision-making

One of the most prominent innovations of AgentCore is the introduction of episode memory, which allows agents to learn from past experiences and improve their decision-making over time. This feature enhances agents’ ability to handle similar tasks more efficiently, reducing the need for manual intervention and optimizing application performance.

As the company itself explains, “For example, an agent schedules transportation to the airport 45 minutes before a flight when traveling alone. Three months later, when traveling to the same destination (this time with children), it automatically programs the pickup two hours in advance, taking into account the challenges of previous family trips. This focused learning approach helps agents make more consistent decisions based on actual performance data instead of relying on predetermined guidelines.”

Photo: AWS