For example, Thumbtack’s API already lets Alexa search for local service providers like plumbers or handymen. But booking and confirming the job still requires going through a website. Nova Act will step in, understanding the page visually, filling out details, and completing the booking.
Nova Act is just one of several specialized AI models Alexa orchestrates to keep conversation natural, execute actions on the web, and validate accuracy. One of the most important breakthroughs will be how it performs under real-world conditions, where customers expect answers in seconds, even as Alexa coordinates multiple models, services, and partners behind the scenes.
The Multi-Model Orchestrator
Delivering speed at scale requires a reimagined foundation and new advances in arbitration and latency. The rearchitected system is built on Amazon Bedrock, a platform for building generative AI applications and agents at production scale. AWS made it model-agnostic so the right model can be applied at each step. Supporting this are what the Alexa team calls “experts”—collections of APIs, workflows, and reasoning tools tailored to domains such as booking a table, managing home services, or controlling smart home devices.
With this architecture, Alexa will be able to arbitrate among providers at runtime, evaluating multiple options for completing a task, surfacing the best fit, and respecting customer preferences. Just as importantly, it will do this instantly. A sophisticated routing system is being developed to minimize latency by matching each request to the fastest and most effective path—balancing speed, accuracy, and reliability without breaking the flow of conversation.
“We have a model for every use case, from complex agentic tasks like coordinating across multiple agents and hundreds of Alexa experts, to really fast natural conversations with minimal latency, to things like navigating a web browser with Nova Act,” Giannangeli says.
AI That Gets Things Done in the Real World