Amazon is putting more weight behind AI agents with a new research and development laboratory in San Francisco. The Amazon AGI SF Lab is focused on systems that can do more than respond to prompts: the goal is to create agents that can perform actions, manage complex workflows, and improve through human feedback.
The move brings together talent connected to Adept and Covariant, while tying the new lab to Amazon’s broader AGI work. It also places Amazon more visibly in a field where Microsoft, Anthropic, and OpenAI are already developing agent-based AI systems.
A new lab for action-oriented AI
The Amazon AGI SF Lab will focus on AI agents that can operate in both digital and physical environments. That matters because the company is not describing agents only as chat tools or workplace assistants. The lab’s stated direction includes systems that can take actions across different kinds of settings.
The work is also aimed at complex workflows. In plain terms, that suggests Amazon is interested in agents that can handle multi-step tasks rather than isolated responses. The source article does not describe specific products from the lab, but it does make clear that action, workflow handling, and learning from human feedback are central to the project.
Human feedback is an important part of that framing. If agents are expected to act, not just answer, then feedback becomes part of how those systems are shaped and corrected. Amazon’s stated focus points toward agents that are meant to become more useful through interaction with people.
Who is leading the Amazon AGI SF Lab
The lab will be led by David Luan, co-founder of AI startup Adept. Adept was among the first startups to develop generative AI-based agents and introduced its ACT-1 model in early 2022.
The lab will also work closely with robotics researcher Pieter Abbeel. He joined Amazon after its "license and hire" deal with Covariant. That detail is notable because the lab’s goals include physical environments, and Abbeel’s work is connected in the source to robotics.
Amazon will initially staff the lab with Adept employees who joined the company through a similar arrangement. That gives the new lab a clear starting point: people who have already worked on generative AI-based agents will be part of Amazon’s internal push.
How this fits with Amazon’s existing AI work
The new San Francisco lab will build on results from Amazon’s broader AGI team. That team recently introduced the new Nova models, which are named in the source as part of the foundation for the lab’s work.
Amazon also already has experience with AI agents through Bedrock and Amazon Q Business. The company sees significant potential in this sector and wants to take a leading position. The new lab gives Amazon another dedicated path for agent research and development, while connecting that effort to existing platforms.
The strategy appears to span more than one use case. The source points to business tools through Bedrock and Amazon Q Business, and to smart assistants through Alexa. That makes the lab relevant both to workplace software and to assistant experiences that could eventually become more capable.
Alexa and the push toward agents
Alexa is also expected to gain more "agent-like" AI capabilities. CEO Andy Jassy has hinted at a version of Alexa that can not only answer questions but also take actions.
That distinction is the core shift behind AI agents. A conventional assistant can provide information. An agent-like assistant is framed as something that can act on behalf of a user, though the source does not provide specific examples of what Alexa would do.
There are still obstacles. According to insiders cited in the source, hardware and reliability issues are causing delays in the new Alexa rollout. That detail is important because it shows the gap between agent ambitions and dependable deployment.
A competitive field is taking shape
Amazon is not working on AI agents alone. Microsoft, Anthropic, and OpenAI are also developing AI agents, according to the source article.
Anthropic has already released an early project called "Computer Use," where Claude 3.5 Sonnet can control a computer. That example shows why the agent category is attracting attention: the focus is shifting from systems that generate content to systems that can interact with tools and environments.
For Amazon, the Amazon AGI SF Lab is another step into that same direction. Its focus on digital and physical environments, complex workflows, human feedback, Adept employees, and the Nova models gives the company a defined research base as AI agents become a more contested part of the AI market.