AWS has held a new edition of re:Invent 2025which, as has already happened on several occasions, has focused on AI and the infrastructure and functions related to it. The number of new features that the company has presented throughout the five days that the event lasts is very numerous, but there are three that stand out above all of them: the frontier agents AI agents, frontier Nova models and their new model customization capabilities to simplify the creation of AI agents.
AWS frontier agents are a new, more sophisticated class of AI agents with three notable characteristics: they are autonomous, scalable, and work independently. This means that if they are given a goal, the agents figure out how to achieve it, that they can perform several tasks at the same time and distribute the work among several agents, in addition to being able to work for hours or even days without human intervention.
The first three frontier agents announced by the company, now available in a trial version, are Kiro Autonomous Agent, AWS Security Agent and AWS DevOps Agent. They are all focused on transforming the software development lifecycle, and can complete complex projects autonomously.
Kiro Autonomous Agent
Kiro Autonomous Agent is a frontier agent for software development that keeps work going independently while the user focuses on priority tasks. This way, he or she can have more time for higher-priority work, since he or she will spend less on background tasks, shortening the path from idea to valuable contributions.
This agent maintains a persistent context between sessions and continually learns from pull requests and feedback. You can handle a variety of tasks, from triaging and prioritizing bugs to improving code coverage, with a single change that spans multiple repositories. You can do prquestions, describe a task, and assign backlog items directly from GitHub.
The agent will independently figure out how to do the work, sharing changes as proposed edits and pull requests, so that control over what is incorporated is maintained at all times. Kiro Autonomous Agent is a shared resource that works together with the entire team, building a collective understanding of the code base, products and standards.
Connects to the team’s repos, pipelines and tools; like Jira, GitHub and Slack; to maintain context as work progresses, adapting to changes or updates. Every code review, every ticket, and every architectural decision fuels the agent’s understanding, making them more useful to the team over time.
AWS Security Agent
AWS Security Agent Helps build secure applications from the ground up in AWS, multicloud, and hybrid environments. Incorporate security expertise throughout the development cycle, proactively reviewing design documents and analyzing pull requests against the organization’s security requirements and common vulnerabilities.
By defining your organization’s security standards once, AWS Security Agent will automatically validate them across applications during each review, helping teams address risks while avoiding generic checklists.
The agent also transforms penetration testing from a slow, manual process to an on-demand capability, in line with the team’s development speed. In this way, penetration testing can now be extended to the entire application portfolio. Additionally, it returns validated findings with remediation code to correct the issues it finds..
If multiple applications are being deployed at once, the number of AWS Security Agents can be scaled to meet demand, so you don’t have to choose between moving fast and maintaining security. By continuously validating security from design to deployment, the agent helps prevent vulnerabilities in the first place.
AWS DevOps Agent
AWS DevOps Agent provides fewer alerts to the team with incident triage always-on, guided resolution, and recommendations to continuously improve application reliability and performance in AWS, multi-cloud, and hybrid environments.
Additionally, you are on call when incidents occur, respond instantly to problems, and use your knowledge of the application and component relationships to find the root cause of the problem. Learn about resources and relationships, ranging from observability tools to runbooks, code repositories, and continuous integration and continuous delivery (CI/CD) pipelines.
Additionally, it maps application resources and correlates telemetry, code, and deployment data to precisely locate root causes of problems and reduce mean time to resolution. You can also move from operating reactively to proactive operational improvement by analyzing patterns in historical incidents with AWS DevOps Agent.
The agent uses those learnings to provide specific recommendations that reinforce four key areas: observability, infrastructure optimization, improving deployment pipelines, and application resiliency. This approach leverages untapped value in data and operational tools and helps teams improve recovery times.
Frontier Nova models
At re:Invent 2025, AWS announced the expansion of its portfolio of Nova models with four new ones, as well as an open training service for companies to create their custom model variants with Nova, and a service for creating AI agents.
The new Nova models, from the Nova 2 range, include Nova 2 Lite, Pro, Sonic and Omni. The first is a rapid reasoning model for everyday workloads capable of processing text, images and videos to generate text. Your users can adjust how much step-by-step reasoning the model performs before responding, balancing depth of intelligence with speed and cost.
It is designed for customer service chatbots, document processing and business automation. It is intended for document processing, extracting information from videos, generating code, delivering informed and accurate responses, and automating multi-step agentic workflows.
Nova 2 Pro is an intelligent reasoning model, capable of processing, in addition to the same font times as the Lite, also voice, with the aim of generating text. It is especially designed for tasks such as agentic coding, long-term planning and sophisticated problem solving, where precision prevails. You can also act as a “teacher” to pass your capabilities to smaller, more efficient “student” models for specific domains and use cases.
Shows strength in multi-document analysis, video reasoning, following complex instructions, solving advanced mathematics, and executing agentic and software engineering tasks. Both Lite and Pro integrate web grounding and code execution capabilities, meaning they can search online for up-to-date information and execute code directly.
Nova 2 Sonic is a speech-to-speech model that unifies text and speech understanding and generation for real-time, human-like conversational AI. Includes expanded multilingual support with expressive voices and a million-token context window for long interactions. In addition, it allows you to switch between voice and text.
The model manages tasks asynchronously, allowing users to follow natural conversations and even change topics, while other actions are completing in the background. It integrates with Amazon Connect, telephone providers or conversational AI frameworks. This makes it ideal for customer service, AI assistants, and interactive voice experiences.
Nova 2 Omni is a multi-modal generation and reasoning model capable of processing text, images, video and voice while generating text and images. Manage up to 750,000 words, hours of audio, long videos and documents of hundreds of pages. In addition, it is capable of analyzing product catalogues, testimonials, brand guides and video libraries at the same time. It shows strength in public benchmarks for multimodal reasoning on documents, images, videos, and audio, and can generate high-quality images.
AWS Nova Forge y Act
In addition to the models, AWS has also introduced Nova Forge, which allows companies to create their optimized variants of Novacombining its proprietary data with Nova’s frontier capabilities. It offers open training, with exclusive access to checkpoints of pre-trained, mid-training, and post-trained Nova models, so customers can mix their proprietary data with data sets chosen by Amazon Nova at every stage of model training.
The result is a customized model that combines Nova’s comprehensive knowledge and reasoning ability with a deep understanding of each company’s specific business. Nova Forge now allows you to create Novellas (custom Novas) starting today, and Forge users will have early access to Nova 2 PRo and Nova 2 Omni.
Nova Forge offers the possibility of training AI models using its own reinforcement learning environments. It also allows you to create smaller, faster models trained with AI-generated examples from larger models. In addition, it gives access to a suite of responsible AI tools, which allows security controls to be implemented.
On the other hand, Nova Act is now available as service on AWS to create and deploy AI agents that can perform actions in web browsers. It is powered by a custom Nova 2 Lite model, and is an avenue to create and manage fleets of agents that automate browser-based tasks.
Simplify model customization with Amazon Bedrock and Amazon SageMaker AI
AWS has provided new capabilities to make advanced model customization available to developers in Amazon Bedrock and Amazon SageMaker AI. Are Reinforcement Fine Tuning (RFT) en Amazon Bedrock and the customizing serverless models in SageMaker AI with reinforcement learning.
RFT on Amazon Bedrock simplifies the model customization process, providing an average of 66% improvement in accuracy compared to the base models. In this way, it allows you to achieve better results with smaller, faster and more profitable models, instead of depending on larger and more expensive models.
Developers choose their base model, link it to their invocation records, or upload a data set. They then choose a reward function based on AI, rules, or a ready-to-use template, while Amazon Bedrock workflows handle the end-to-end fine-tuning. At launch, RFT on Amazon Bedrock will be compatible with Amazon Nova 2 Lite, with support for more models coming soon.
SageMaker AI now has more serverless model customization capabilities, making model customization possible in days. Developers who use it can choose between two experiences: an agentic one, available in a trial version, which uses an agent to guide developers through the model customization process; and a self-directed approach, for those who prefer to be in control of the process.
With the first experience, developers describe what they need in natural language, and the agent goes through the entire customization process. But with the second, developers looking for granular control and flexibility will see infrastructure management eliminated, and will have the right tools to choose the customization technique and adjust parameters.
Both options provide access to advanced personalization techniques, such as reinforcement learning from AI feedback, with verifiable rewards, supervised fine-tuning, and direct preference optimization.
