Rebuilding Alexa: How Amazon is mixing models, agents and browser-use for smarter AI

Amazon Betting on agent interoperability and model mixing, we make the new Alexa Voice Assistant more effective and retool the flagship voice assistant with agent functionality and browser usage tasks.

This new Alexa has been rebranded as Alexa+, and Amazon has highlighted this version “more things.” For example, you can actively tell users whether a new book from your favorite author is available, your favorite artist is in town, or even if they would offer to buy tickets. For Alexa+ reasons, tap “Expert” on various knowledge bases to answer user questions or answer “Where is the closest pizza place to the office? Does my colleague like it? – If you think they’ll do you want to make a reservation.”

In other words, ALEXA+ blends AI agents, computer usage capabilities, and knowledge to learn from the larger Amazon ecosystem, and wants Amazon to be a more capable and smarter home voice assistant.

Alexa+ is currently running on Amazon’s NOVA model and models from Humanity. However, Daniel Rausch, VP of Amazon’s Alexa and Echo, told VentureBeat that the device remains a “model agnostic” and that the company can introduce other models (at least models available on Amazon Bedrock) to find the best one to accomplish the task.

“[It’s about] Choose the right integration to complete the task, get the right kind of instructions, know what you need to actually complete the task, then adjust the whole thing,” said Rausch. “The big thing to understand about that is that Alexa continues to evolve with the best models available anywhere in Bedrock.”

What is model mixing?

Model mixing or model routing allows businesses and others to choose the right AI model to tap on a per-query basis. Developers are increasingly relying on model mixing to reduce costs. After all, not every prompt needs to be answered with an inference model. Some models perform a better specific task.

Amazon’s cloud and AI units, awshas long been a supporter of model mixing. Recently, we have announced a feature called Bedrock’s intelligent prompt routing. This will point the prompt to the optimal model and model size to resolve the query.

And it could be working. “You can say that you can’t say which model you’re using for a particular response from Alexa for a particular task,” says Rausch.

Agent Interoperability and Orchestration

Rausch said Alexa+ brings together agents in three different ways. The first is the traditional API. The second is to deploy agents that can navigate websites and apps such as using Anthropic’s computers. The third is connecting the agent to other agents.

“But at the heart of it all, adjusting across all these different kinds of experiences is a very capable and cutting-edge LLMS at these baselines,” Rausch said.

He added that if a third-party application already has its own agent, the agent can talk to agents working within Alexa+, even if the external agent is built using a different model.

Rausch highlighted that the Alexa team is using Bedrock tools and technologies, including new multi-agent orchestration tools.

Humanity’s CPO Mike Krieger told VentureBeat that previous versions of Claude could not achieve what Alexa+ wanted.

“Really funny” why now? “Of course, the moment is clear in the demo as the models are better,” Krieger said. “But if you try to do this with a 3.0 sonnet or 3.0 level model, I think you’ll struggle in many ways to use many different tools at once.”

Neither Rausch nor Krieger confirms the specific human model that Amazon used to build Alexa+, but it is worth pointing out that humanity released Claude 3.7 Sonnet on Monday, and is available on Bedrock.

Large investment in AI

The first brushes of many users with AI have gone through AI voice assistants such as Alexa, Google Home and even Apple’s Siri. They allow people to outsource some tasks, such as turning on the lights. I don’t own an Alexa or Google Home device, but I recently learned how convenient it is when staying in a hotel. I was able to tell Alexa to turn off the alarm, turn on the lights, and open the curtains while I was under the cover.

However, while Alexa, Google Home Devices and Siri have become ubiquitous in people’s lives, they have begun to show an age when generative AI has become popular. Suddenly, people wanted more real-time answers from AI assistants, demanding smarter task resolution, such as adding multiple meetings to the calendar without the need for many prompts.

Amazon has admitted that the rise of AI generals, particularly agents, has enabled Alexa to ultimately fulfill that possibility.

“Up to this moment, we were limited by technology as to what Alexa is,” said Panos Panay, Amazon’s device and services SVP, during the demonstration.

Rausch said he hopes that Alexa+ will continue to improve, add new models and make more people comfortable with the technology being able to do.