Amazon is giving AI agents a more visible place in its research agenda with a new San Francisco lab focused on systems that can act across software and, potentially, the physical world.
The Amazon AGI SF Lab will work on what Amazon calls “foundational” capabilities for AI agents. The goal is not simply to build chat systems, but to develop agents that can use computers, web browsers and code interpreters to handle more complex workflows.
A new lab built around action-taking AI
The lab will be led by David Luan, co-founder of AI startup Adept. Amazon says the group will seek to build agents that can “take actions in the digital and physical worlds” and “handle complex workflows” using common computing tools.
That framing matters because AI agents are being positioned as software that can move beyond answering questions. In this vision, an agent can understand a task, interact with tools, make progress through several steps and adjust when the path changes.
In a post written by Luan and Pieter Abbeel, the pair said the work would build on Amazon’s broader AGI team. Abbeel, a robotics researcher who joined Amazon after the company’s “license and hire” deal with Covariant, will work “closely” with Luan and the AGI SF Lab, according to an Amazon spokesperson cited by TechCrunch.
Luan and Abbeel described the lab’s early direction this way: “Our initial focus is on several key research bets that will enable AI agents to perform real-world actions, learn from human feedback, self-course-correct, and infer our goals.”
Why Adept is central to the effort
The new lab is closely tied to Amazon’s earlier arrangement with Adept. In June, Adept agreed to license its technology to Amazon, and Luan and portions of Adept’s team joined the company.
Adept was founded two years ago around a direct and ambitious idea: create an AI model that can perform actions on any software tool using natural language. At a high level, the company’s vision was an “AI teammate” trained to use a wide range of software tools and APIs.
Amazon is now using that talent and technology base as part of a broader AGI effort. Luan was working under Rohit Prasad, the former head of Alexa, and will continue to do so. Prasad leads an AGI team focused on large language models.
The Amazon AGI SF Lab will be seeded by Adept employees. Amazon also says it is looking to hire a few “dozen” additional researchers, including people from fields such as quantitative finance, physics and math.
The competitive race around agentic AI
Amazon is not pursuing AI agents in isolation. The broader market for “agentic” AI has become crowded, with startups and large AI companies working on systems that can complete tasks with less direct user control.
According to Emergen Research, the “agentic” AI sector could be worth $31 billion by the end of the year. A Capgemini poll also found that Eighty-two percent of orgs say they plan to integrate AI agents within three years, drawn by possible efficiency gains.
The source article points to several companies working in this area, including startups such as Orby, Emergence and Rabbit. OpenAI and other major AI players are also developing related products that can complete tasks largely on their own.
Anthropic earlier this year released its own take on the technology. Google is reportedly working on AI agents that could make purchases for users, including booking flights and hotels.
For Amazon, the San Francisco lab signals a more serious research push in a field where it has already made several moves. In July, the company announced conversational agents for its Bedrock AI development platform. Just last week, it brought agents to Amazon Q Business, its assistant platform for enterprise customers and devs.
How this connects to Alexa and Amazon’s wider AI plans
Amazon CEO Andy Jassy has also hinted at a more “agentic” Alexa. The idea, as described in the source article, is an Alexa that can do more than answer questions and can take actions as well.
That direction aligns with the lab’s stated focus. If agents are expected to perform real-world actions, learn from human feedback, self-correct and infer goals, then the challenge is broader than a single product interface. It reaches into how AI systems understand intent, operate tools and recover from mistakes.
The regulatory backdrop is also part of the story. Amazon’s Adept arrangement resembled Microsoft’s deal with AI startup Inflection in May. Both deals have drawn regulatory scrutiny as policymakers stateside and abroad examine whether large technology companies are smothering AI rivals.
The Amazon AGI SF Lab therefore sits at the intersection of research, product strategy and competition. It gives Amazon a dedicated team in San Francisco for AI agents, led by a founder whose startup was built around software-operating AI. It also places Amazon more clearly in the same race as startups and major AI labs trying to turn agentic AI from a promise into a practical layer of computing.