Amazon makes it simpler to build efficient AI agents
Key takeaways
- New Amazon Bedrock and Amazon SageMaker AI capabilities give customers access to advanced techniques for tailoring models.
- Reinforcement Fine Tuning in Amazon Bedrock makes it easier to adapt models to unique use cases and boost accuracy.
- Amazon SageMaker AI speeds up advanced model customization from months to days, helping AI development accelerate and new solutions reach market faster.
Efficiency has become a central challenge for organizations deploying AI. While creating AI applications has become easier, running them at scale remains costly and resource-intensive. This challenge is especially pronounced for AI agents, which can demand higher inference when they reason through problems, leverage multiple tools, and coordinate across systems. Many companies default to the largest, most powerful models to power their agents, yet a large portion of an agent’s time goes to routine tasks—like checking calendars and searching documents—that don’t need advanced intelligence. The result is unnecessary costs, slower responses, and wasted resources.
The path forward is customization: tailoring smaller, specialized models to handle the frequent tasks that agents perform, delivering faster, more accurate responses at a lower cost. Until now, advanced customization techniques—such as reinforcement learning—required deep machine learning expertise, substantial infrastructure, and months of development time.
Today, a new era begins with 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 than base models. By making these techniques more approachable for developers, organizations of all sizes can build custom agents for any business need.
RFT made approachable for everyday developers with Amazon Bedrock
Traditional customization methods can be a barrier to building truly efficient models. Reinforcement learning trains a model using feedback from humans or another model, reinforcing good behavior and correcting errors. It’s especially effective for reasoning and complex workflows because it rewards solid processes, not just correct answers. However, reinforcement learning demands a complex training pipeline, massive compute, and access to expensive human feedback or a powerful evaluation model.
RFT on Amazon Bedrock lowers the barrier, making the technique accessible to any developer at any organization. Amazon Bedrock is a fully managed AI platform that gives customers access to high-performing foundation models from leading AI providers, plus tools to build agents and generative AI applications with security, privacy, and responsible AI features. RFT on Bedrock delivers about a 66% accuracy improvement on average over base models, enabling better results with smaller, faster, and more cost-effective models instead of relying on larger, expensive options.
The workflow is straightforward. Developers pick a base model, point it at invocation logs (the AI’s history), or upload a dataset. Then they choose a reward function—AI-based, rule-based, or a ready-made template. Automated Bedrock workflows handle the fine-tuning end-to-end. No PhD in machine learning required—just a clear sense of what good results look like for the business. At launch, RFT in Bedrock will support the Amazon Nova 2 Lite model, with compatibility for additional models coming soon.
Customers like Salesforce and Weni by VTEX have seen increased accuracy and efficiency using RFT in Bedrock. Phil Mui, SVP of Software Engineering, Agentforce at Salesforce, said, “AWS’s benchmarking with Bedrock’s Reinforcement Fine Tuning shows promising results, with up to 73% improvement in accuracy over the base model for our specific business requirements. We expect to use 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 upholding Salesforce’s high standards for quality and safety.”
SageMaker AI speeds up model customization from months to days
Teams that want more control over the AI workflow can turn to SageMaker AI. Developers opt for SageMaker AI when they need full control to build, train, and deploy the most capable models at scale.
Since its inception in 2017, SageMaker AI has streamlined the AI development process. Yet as organizations adopt more advanced customization techniques, there’s a growing demand for a smoother experience that eliminates roadblocks—like infrastructure management and synthetic data generation—so focus can stay on delivering better outcomes for customers. To address this, SageMaker AI now supports serverless model customization capabilities, enabling model customization in just days.
Two experiences are available: an agentic experience (launched in preview) that uses an agent to guide developers through the customization process, and a self-guided approach for those who prefer full control. In the agentic experience, developers describe their needs in natural language and the agent manages the entire customization process, from generating synthetic data to evaluation. Those seeking granular control can choose the self-guided experience. This approach eliminates infrastructure management while providing the right tooling to select a customization technique and adjust parameters.
Across either path, developers gain access to 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 are compatible with Amazon Nova and popular open-weight models such as Llama, Qwen, DeepSeek, and GPT-OSS, offering a broad spectrum of options to match the right model to each use case.
Companies like Collinear AI, Robin AI, and Vody are already leveraging SageMaker AI’s new capabilities to simplify model customization. For instance, Collinear AI, an AI improvement platform for enterprise GenAI, saved weeks by adopting SageMaker AI. Soumyadeep Bakshi, co-founder of Collinear AI, explained, “Fine-tuning AI models is crucial for creating high-fidelity simulations, but it used to require stitching together multiple systems for training, evaluation, and deployment. Now, with Amazon SageMaker AI’s new serverless model customization, we have a unified approach that cuts experimentation cycles from weeks to days. This end-to-end serverless tooling lets us focus on building better training data and simulations for our customers, not on 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.
)
