Amazon Web Services (AWS) today extended the reach of its platforms for building and deploying artificial intelligence (AI) applications and agents to make it simpler to customize and finetune models.
Announced at the AWS re:Invent 2025 conference, these additions to the Amazon Bedrock service for accessing AI models and the managed Amazon SageMaker service promise to reduce the total cost of building and deploying these classes of applications while improving the quality of the outputs generated.
Swami Sivasubramanian, vice president for agentic AI at AWS, told conference attendees it’s too challenging to customize AI models to achieve that goal. The addition of these capabilities makes it easier to apply reinforcement learning to existing models that reinforce good behavior in a way that also serves to correct bad behavior. Data science teams can select a base AI model, expose it to invocation logs and additional data sets, with automated workflows then handling the fine-tuning process to improve accuracy by as much as 66% on average, said Sivasubramanian.
At launch, these capabilities are being made available for the Amazon Nova 2 Lite model, with support for additional models planned.
The serverless model customization capabilities being added to Amazon SageMaker, meanwhile, come in two forms. AWS is previewing an AI agent that guides development teams through the model customization process after they describe their intent in natural language, or alternatively they can take advantage of a more self-guided option. These SageMaker AI capabilities will work with Amazon Nova, Llama, Qwen, DeepSeek, and GPT-OSS models.
At the same time, AWS has added a Checkpointless training capability to SageMaker HyperPod that makes it simpler to recover from faults that might occur by making it easier to begin again from a previous point in time.
AWS is also previewing TypeScript support for the open source Strands Agents software development kit (SDK) that it launched earlier this year to provide development teams with modeling tools to build AI agents, along with Edge Device support for Strands Agents that can be used to build applications.
Strands Agents is part of the AWS Bedrock AgentCore portfolio of tools that AWS provides for building and deploying AI agents. Since being launched earlier this year in preview, AWS today revealed there have been more than two million downloads, with organizations such as Cohere Health, Cox Automotive, PGA TOUR, Thomson Reuters, Snorkel, and Swisscom using the framework.
These additions extend a portfolio of AI tools and platforms that AWS extended earlier this week as part of an effort to reduce the cost of AI.
Nick Patience, vice president and AI practice lead at Futurum Group, said the addition of, for example, Trainium3, Ultraserver platforms and the AI Factories service is making it the least expensive place to run AI workloads in production environments.
AWS is also at the same time breaking down the model training process to make it easier to build their own models, he added.
The challenge and opportunity, of course, now is determining which model to deploy when and where based on not just the use case but also the total cost.
