Each edition of the AWS re:Invent conference in Las Vegas sets the azimuth that the entire cloud industry follows over the following months. However, if in previous years the theme was that humans were ‘assisted’ by artificial intelligence – as we know it in the form of Copilots and chat assistants – 2025 has brought a fundamental change in narrative. AWS stops talking about AI that waits for your commands. It is starting to talk about AI that acts on its own. The slogan of this year’s edition is ‘autonomy’, and the new class of Boundary Agents unveiled and the Nova family of models signal that we are entering an era where software stops being just a tool and becomes an autonomous performer.
The number of innovations announced over the five days of the conference is overwhelming, but once the marketing hype is sifted away, three key innovations remain in the battleground that have the potential to redefine the software development lifecycle (SDLC) and IT operations. These are: a new category of autonomous frontier agents (Frontier AI Agents), the powerful and diverse Frontier Nova family of models and, just as importantly, the democratisation of the process of adapting these models to specific business needs.
Border agents: The end of AI micromanagement
Until now, interacting with generative AI has been akin to working with an intern who needs to be watched every step of the way. AWS has decided to change this by introducing the concept of Frontier Agents. This is a new, more sophisticated class of software that is distinguished by three characteristics: autonomy, scalability and independence.
In practice, this means a paradigm shift from ‘chat-based’ to ‘goal-based’. The user no longer needs to lead the agent by the hand through the steps of the process. It is enough to define the end goal. These agents are able to break down a complex task into subsections, distribute the work among themselves and, most importantly, work in the background for hours or even days without any human intervention. It is this persistence of operation and ability to maintain context over a long period of time that sets them apart from legacy solutions.
On the frontline of this revolution is the Kiro Autonomous Agent. It is a solution that has the potential to become every programmer’s ‘digital colleague’. Unlike standard code assistants, Kiro is not limited to syntax prompting in the IDE. It is a shared resource that collaborates with the entire team. It integrates with repositories, CI/CD pipelines and tools such as Jira, GitHub and Slack. In this way, it builds an ‘understanding’ of the company’s code base, products and standards.
Kiro can independently manage backlogs, prioritise bugs and even propose code changes involving multiple repositories simultaneously. Crucially, the agent learns from historical pull requests and feedback, becoming an increasingly competent team member over time. Developers can delegate tasks to it (‘fix this bug and check for other services’), while they themselves focus on higher-priority conceptual work. Final control remains in the hands of the human (through PR approval), but the ‘dirty work’ is done automatically.
DevSecOps on autopilot
AWS’ vision for security and operational maintenance (DevOps) is equally impressive. The introduction of AWS Security Agent and AWS DevOps Agent is a response to the chronic shortage of specialists in these areas and the increasing complexity of distributed systems.
AWS Security Agent has the ambition to transform the approach to security from reactive to proactive. Instead of relying on static checklists run at the end of the process, this agent accompanies the application lifecycle from the very beginning. It analyses project documentation and code for organisation-specific security vulnerabilities. Perhaps the biggest innovation, however, is the reduction of penetration testing to an on-demand service. Instead of slow, manual audits, the agent can perform penetration testing scalable to the entire application portfolio, providing ready-made pieces of remediation code. This solves the problem of choosing between ‘rapid deployment’ and ‘secure deployment’.
AWS DevOps Agent, on the other hand, targets the reduction of so-called ‘alert fatigue’ (alert fatigue). In modern multi-cloud and hybrid environments, finding the needle in the haystack (i.e. the cause of the failure in gigabytes of logs) is a challenge. This agent not only monitors the system 24/7, but at the time of an incident, it independently correlates telemetry data, code and deployment history to pinpoint the root cause of the problem (Root Cause Analysis). What’s more, it works proactively – analysing past patterns to suggest optimisations to the infrastructure before a failure even occurs.
Nova 2: Tailor-made models, not ‘one for all’
The engine driving these autonomous entities is the new Frontier Nova 2 family of models. AWS is clearly moving away from the race for one giant model that can do everything. Instead, the company is offering a portfolio of specialised tools, selected according to the balance between intelligence, speed and cost.
The new family includes four models:
1 Nova 2 Lite: A fast reasoning model designed for everyday business workloads, customer service chatbots or document processing. Its unique feature is that the user can customise the depth of reasoning (step-by-step reasoning). This allows the user to precisely control the cost and speed of the response depending on the complexity of the query.
2 Nova 2 Pro: This is the intellectual ‘heavy artillery’. As well as text and voice processing, this model specialises in high-precision tasks: agent coding, advanced mathematics and long-term planning. An interesting feature is its role as a ‘teacher’ – Nova 2 Pro can be used to distil knowledge and train smaller, more efficient models for specific domains.
3 Nova 2 Sonic: A true breakthrough in voice interaction. It is a model that unifies the understanding and generation of text and speech, enabling real-time conversation with minimal latency. It supports expressive voices and allows you to interrupt speech or change the subject on the fly, making it an ideal candidate for the next generation of voice assistants.
4 Nova 2 Omni: A multimodal giant that blurs the boundaries between data formats. It can process text, images, video and voice simultaneously. Its capabilities are impressive – from analysing hours of video and hundreds of pages of documentation, to simultaneously processing product catalogues and multimedia libraries.
Both the Lite and Pro models have built-in ‘grounding’ (grounding) capability in the network and code execution, eliminating the problem of hallucinations based on outdated data.
Democratisation and personalisation: Nova Forge and Bedrock
Technology is one thing, but implementing it in a business is another. AWS understands very well that the greatest value for companies is not a ‘raw’ model, but a model that knows the specifics of their business. That is why, alongside the models themselves, powerful tools for customising them are presented.
AWS Nova Forge is an environment that allows companies to create so-called ‘Novellas’ (i.e. custom variants of Nova models). With open training services, customers can combine their proprietary data with the powerful capabilities of Frontier’s models at every stage – from pre-training to post-training. This allows for a model that combines general world knowledge with a detailed understanding of the company’s internal processes, with full control over data security. Users can also use their own Reinforcement Learning environments.
Complementing this offering are new features in Amazon Bedrock and Amazon SageMaker AI. The introduction of Reinforcement Fine Tuning (RFT) in Bedrock simplifies the process of fine-tuning models, offering an average of up to 66% improvement in accuracy over baseline models. This is a key insight for CFOs – instead of investing in gigantic, inferentially expensive models, companies can achieve better results by using smaller, cheaper and faster models that have been finely-tuned for a specific task.
Amazon SageMaker AI, on the other hand, introduces serverless model customisation, which drastically reduces implementation time – from weeks to days. Developers are given a choice: they can use an agent to guide them through the process in natural language, or take full control of the parameters if the specific needs of the project require it.
Are we ready for autonomy?
This year’s re:Invent is more than just the launch of new products. It is a manifesto for a new philosophy of working with technology. AWS’s thesis is that the future of IT belongs not to people who click the keyboard faster thanks to AI prompts, but to people who can manage a fleet of autonomous agents doing the work for them.
Tools such as Kiro, Nova Act (which allows agents to perform actions in browsers) or specialised Nova 2 models, are creating an ecosystem where the barrier to entry for advanced software development is drastically decreasing and operational efficiency is increasing. However, an open question remains: how soon will organisations be ready to trust the autonomy of agents? Putting control of the code repository or security testing in the ‘hands’ of AI requires not only a technological change, but above all a mental one. One thing is certain – after re:Invent 2025, the AI arms race has moved to a whole new level.
