AWS has lowered the entry barrier for AI customization automation through reinforcement learning.

Amazon Web Services (AWS) has released new features aimed at improving artificial intelligence (AI) model efficiency and reducing training costs. At the core are the reinforcement learning-based fine-tuning option “RFT” and serverless customization capabilities. These features are designed to enable developers to make user-specific improvements to AI models with minimal machine learning expertise.

On the 3rd (local time), AWS announced at its annual “AWS re:Invent 2025” conference in Las Vegas that these features will be applied to Amazon Bedrock and SageMaker AI. Amazon Bedrock is a platform for building generative AI capabilities based on “foundation models” from leading AI companies. The RFT update paves the way for enterprises to optimize AI agents without needing powerful machine learning infrastructure.

Enterprises typically use the highest-performing large language models (LLM) for AI agents, which results in excessive inference processing requirements. Even for repetitive tasks such as calendar confirmation or document retrieval, there are inefficiencies caused by excessive resource usage. AWS believes that the newly introduced reinforcement learning-based customization features can address these issues. In short, this is a structure that ensures sufficient efficiency with less computation.

Previously, introducing reinforcement learning came with extremely high technical barriers. Tasks such as preparing training data, collecting feedback, and building high-performance computing infrastructure could take months. However, the RFT feature on Amazon Bedrock allows developers to select the desired model, upload user interaction logs or training data, and specify a reward function, after which model tuning is performed automatically. AWS explains that this process can be executed without machine learning professionals and that simply having “an idea of what a good result is” is sufficient.

Initially, this feature will exclusively support Amazon’s own model, Nova 2 Lite, but there are plans to expand it to dozens of models in the future. Amazon SageMaker AI will also add similar functionality in a serverless form. SageMaker is a platform that enables enterprises to design and deploy their own AI models, and is expected to provide a more easily integrated reinforcement learning “agentized” environment.

In agentized mode, users can enter their requirements in natural language, and the AI agent will guide the entire process from data generation to model evaluation. At the same time, it also provides self-guided methods for advanced developers, broadening the range of user options. AWS explains that this feature can also simultaneously apply various reinforcement learning techniques such as feedback learning, verifiable reward-based learning, and supervised learning-based tuning. The feature is compatible not only with Nova, but also with models such as Llama, Qwen, DeepSeek, and GPT-OSS.

Meanwhile, AWS also announced the introduction of “checkpointless training” in SageMaker HyperPod. Previously, if an error occurred during training, recovery could take over ten minutes, but now the state can be restored within minutes without customer intervention. This is achieved by saving the model state in real time across the entire cluster.

In addition, AWS has ported the open-source AI agent framework “Strands Agents” to the TypeScript language. TypeScript is more stable and less error-prone than JavaScript, and is expected to provide a more reliable environment for agent development.

This release comes amid a trend of competitors in the generative AI market strengthening their customization features. Google (GOOGL), Microsoft (MSFT), and others are also accelerating the rollout of similar features, and it is expected that the environment for users to easily build optimal AI models themselves will develop even faster. Such technological evolution appears to act as a catalyst, driving AI deeper into practical enterprise environments.