Amazon makes it easier to build efficient AI agents

Key takeaways

New Amazon Bedrock and Amazon SageMaker AI capabilities give customers access to advanced techniques for model customization.
Reinforcement Fine Tuning in Amazon Bedrock makes it easier to tailor models to unique cases and improve accuracy.
Amazon SageMaker AI cuts advanced model customization workflows from months to days, accelerating AI development and bringing new solutions to market faster.

Efficiency has emerged as a critical challenge for organizations deploying AI. While building AI applications has become easier, running them at scale remains expensive and resource-intensive. This challenge is particularly acute for AI agents, which can have higher inference demands as they reason through problems, leverage a variety of tools, and coordinate across multiple systems. Most companies opt for the largest, most capable models to power their agents, but a significant amount of an agent’s time is spent doing routine tasks, like checking calendars and searching documents, that don't require advanced intelligence. The result? Unnecessary costs, slower responses, and wasted resources.

Laptop displaying SageMaker AI model selection interface with logos of Amazon, Meta, Qwen, and OpenAI

The solution lies in customization: tailoring smaller, specialized models to handle the work agents do most often to deliver faster, more accurate responses at lower costs. But until now, advanced customization techniques like reinforcement learning required deep machine learning expertise, extensive infrastructure, and months of development time.

Today, we announced new Amazon Bedrock and Amazon SageMaker AI capabilities that make advanced model customization accessible to developers at any organization. Reinforcement Fine Tuning (RFT) in Amazon Bedrock and serverless model customization in Amazon SageMaker AI with reinforcement learning simplify the process of creating efficient AI that's fast, cost-effective, and more accurate compared to base models. By making these techniques more accessible for our customers’ developers, we’re making it easier for organizations of all sizes to build custom agents for any business need.

RFT made easy for everyday developers with Amazon Bedrock

Difficult customization techniques present a roadblock for building custom, efficient models. Reinforcement learning, for example, trains a model using feedback from either humans or another model. Good behavior gets reinforced, while bad behavior gets corrected. It’s particularly good for reasoning and complex workflows because it rewards good processes, not just good answers. However, reinforcement learning requires a complex training pipeline, massive compute, and access to expensive human feedback or a powerful AI model to evaluate every response.

AWS interface for creating reinforcement fine-tuning job with Amazon Nova 2 model

RFT on Amazon Bedrock simplifies the model customization process, opening the technique to any developer at any organization. Amazon Bedrock is a fully managed AI platform giving customers access to high-performing foundation models from leading AI companies, along with capabilities to build agents and generative AI applications with features for security, privacy, and responsible AI. RFT on Amazon Bedrock delivers 66% accuracy gains on average over base models, helping you get better results with smaller, faster, more cost-effective models instead of relying on larger, expensive ones.

The process is simple. Developers select their base model, point it at their invocation logs (in other words, the AI’s history), or upload a dataset. Then, they choose a reward function—AI-based, rule-based, or a ready-to-use template. Automated workflows in Amazon Bedrock handle the fine-tuning process end-to-end. No PhD in machine learning required—only a clear sense of what good results look like for the business. At launch, RFT in Amazon Bedrock will support the Amazon Nova 2 Lite model. Compatibility with additional models is coming soon.

Customers like Salesforce and Weni by VTEX have seen increased accuracy and efficiency using RFT in Amazon Bedrock. Phil Mui, SVP of Software Engineering, Agentforce at Salesforce, said, “AWS’s benchmarking with Amazon Bedrock’s Reinforcement Fine Tuning shows promising results, demonstrating up to 73% improvement over base model in accuracy for our specific business requirements. We anticipate leveraging RFT to enhance and extend what we already achieve with supervised fine-tuning, enabling us to deliver even more precise and customized AI solutions for our customers. This approach complements our existing AI development workflow while maintaining Salesforce’s high standards for quality and safety.”

Amazon SageMaker AI accelerates model customization from months to days

Laptop displaying Amazon Bedrock custom models interface

Teams that need more control over the AI workflow can turn to Amazon SageMaker AI. AI developers choose SageMaker AI for customization because it gives them full control to build, train, and deploy the most capable models at scale.

Since launching in 2017, SageMaker AI has made the AI development workflow faster and more efficient. However, as organizations look to use more advanced customization techniques, they want more seamless experiences that remove roadblocks that take months of work—like infrastructure management and generating synthetic data—so they can focus on developing better outcomes for customers. That’s why SageMaker AI now supports new serverless model customization capabilities, making model customization possible in just days.

There are two experiences to choose from: an agentic experience, launching in preview, that uses an agent to guide developers through the model customization process, or a self-guided approach for those who like to be in the driver's seat. With the agentic experience, developers describe what they need in natural language and then the agent walks through the entire customization process, from generating synthetic data to evaluation. Developers who want granular control and flexibility can choose the self-guided experience. This eliminates infrastructure management while providing the right tooling to select a customization technique and the ability to tweak the parameters.

With either option, developers can access advanced customization techniques like Reinforcement Learning from AI feedback, Reinforcement Learning with Verifiable Rewards, Supervised Fine-Tuning, and Direct Preference Optimization. The new SageMaker AI capabilities will work with Amazon Nova and popular open weight models like Llama, Qwen, DeepSeek, and GPT-OSS, giving customers a wide range of options to match the right model to their use case.

Collinear AI, Robin AI, and Vody are just a few of the customers that have started simplifying model customization with SageMaker AI’s new capabilities. For example, Collinear AI, an AI improvement platform built for enterprise genAI, saved weeks using SageMaker AI. Soumyadeep Bakshi, co-founder, Collinear AI, said, “Fine-tuning AI models is critical to creating high-fidelity simulations, and it used to require stitching together different systems for training, evaluation, and deployment. Now with Amazon SageMaker AI's new serverless model customization capability, we have a unified way that empowers us to cut our experimentation cycles from weeks to days. This end-to-end serverless tooling helps us focus on what matters: building better training data and simulations for our customers, not maintaining infrastructure or juggling disparate platforms.”

For more details about these capabilities, visit:

Get the latest news from AWS re:Invent, including all things agentic and generative AI, product and service announcements, and more.