After unveiling a major expansion of the Nova foundation model platform to encompass frontier artificial intelligence reasoning on Tuesday, Amazon Web Services Inc. today shifted its focus to the tools and platforms supporting agents.

A key element for AWS is ease of use, a recognition that AI’s inherent complexity can be an obstacle to enterprise deployment and affect confidence in an agent’s ability to handle critical tasks.

“Agents give you the freedom to build without limits,” Swami Sivasubramanian (pictured), vice president of AWS Agentic AI, said in his keynote remarks at AWS re:Invent. “Who can build is rapidly changing. How quickly you can build is also changing. Yet building and scaling these agentic systems are harder than the problems they are trying to solve.”

Updates for popular Strands Agents

The company announced releases designed to streamline the development process through enhancements to tools such as Amazon Strands Agents.

Strands is an open-source SDK launched by AWS in May. It takes a model-driven approach to building and running AI agents using just a few lines of code. AWS announced today that it’s bringing Strands to the TypeScript programming language, generally considered to be more resistant to errors and bugs.

“In just the last few months, Strands has been downloaded more than 5 million times,” Sivasubramanian noted. “You want the ability to rapidly deploy agents at scale.”

Along with the addition of TypeScript, edge device support for Strands is now generally available. AWS also announced feature in preview that will allow developers to systematically validate agent behavior, measure improvements, and deploy with confidence during development cycles.

The latest releases were designed to address the proliferation of new uses for the tool as downloads continue to rise, according to Clare Liguori, senior principal software engineer for AWS Agentic AI.

“We heard from customers that they needed to be able to guide agent behavior in production,” Liguori told SiliconANGLE in an exclusive interview during the conference. “What I’ve learned in my 11 years at AWS is that customers will always use your products in ways you did not expect.”

Addressing model efficiency

Today’s announcements from AWS also focused on an issue which has hindered widespread agentic deployment: model accuracy and efficiency. Organizations are reluctant to spend significant amounts of time and money on customizing models so that agents can perform seemingly routine tasks.

“Today’s models are not the most efficient,” Sivasubramanian said. “The key to success here is quality over quantity. What if we removed the complexity and cost while still giving you access to these advanced fine-tuning techniques?”

The answer, according to AWS, is to make it easier for developers to customize models by using reinforcement learning. Today’s announcement of a Reinforcement Fine Tuning feature in Amazon Bedrock, along with serverless model customization capabilities in Amazon SageMaker AI, are designed to allow companies to create agents using advanced large language models without the need for massive amounts of processing power.

“It’s a way for customers to build models that perform better for them over time, without additional expertise required,” Sivasubramanian told the re:Invent gathering. “You can choose the right approach based on your comfort level.”

Along these lines, today’s announcement of updates for Amazon SageMaker HyperPod was also structured to support more reliable and efficient model training experiences. A new “checkpointless” training feature can preserve the model training state across distributed clusters, an important consideration as GPU usage and cluster size have grown exponentially.

“This is a paradigm shift in model training,” Sivasubramanian said. “Now you can recover training from faults in minutes across thousands of AI accelerators.”

Agents for outer space

It’s one thing to build an agent to check a company calendar and find documents. It’s quite another to employ agents for building a rocket that can be placed into orbit around the earth.

Today’s keynote at re:Invent included an appearance by William Brennan, vice president of enterprise technology at Blue Origin LLC, who described the aeronautics firm’s adoption of agentic AI. Executives at the company have described the company’s vision to scale up AI and lower the cost of access to space. That has included the development of an internal AI platform called BlueGPT, along with agentic workflow systems to design and manufacture space systems.

“Agentic AI has exploded at Blue Origin,” Brennan told the re:Invent audience. “Everyone at Blue is expected to build and collaborate with AI agents. We believe in a world where we can agentically design an entire rocket.”

The announcements this week from AWS reinforced the company’s key message that it’s heavily focused on becoming the central resource for building and running AI agents. For that strategy to be successful, organizations will have to be confident that autonomous technology can be trusted, a reality that Sivasubramanian fully acknowledged.

“We need to be able to trust that agents will perform as they are expected,” Sivasubramanian said. “The future of agentic AI is not on agents that can do everything, it’s on agents we can rely on to do everything.”

Photo: Robert Hof/SiliconANGLE