Amazon launches San Francisco lab to build action-taking AI agents

The frontier of AI is rapidly shifting from models that just talk to agents that can actually do things on your screen and in the real world.

Amazon made its biggest move yet in that direction on Monday, announcing a brand-new San Francisco R&D hub called the Amazon AGI SF Lab. The dedicated facility will focus entirely on building foundational capabilities for AI agents that can use computers, browse the web, execute code, and tackle complex end-to-end workflows.

The Adept crew takes the reins

The new lab is headed by David Luan, the co-founder and former chief of Adept, the buzzy agent startup Amazon quasi-acquired back in June through an unconventional license-and-hire arrangement.

Luan and his former Adept colleagues will form the core of the new hub, and Amazon is already looking to bring on several dozen more researchers with backgrounds in physics, math, and quantitative finance.

Interestingly, the initiative also pulls in UC Berkeley robotics pioneer Pieter Abbeel, who joined Amazon following a similar deal with robot manipulation startup Covariant. Amazon confirmed Abbeel will work closely with Luan and the SF lab.

According to Luan and Abbeel, their initial research bets center on four tricky frontiers: getting agents to perform physical and digital real-world actions, learn from human feedback, correct their own courses mid-task, and accurately infer human intentions.

Inside Amazon, Luan still reports directly to Rohit Prasad, the former Alexa chief who now steers Amazon's broader AGI and large language model efforts.

Racing into a $31B battleground

The timing is no accident. Market tracker Emergen Research expects the agentic AI market to reach $31 billion by the end of the year, while a recent Capgemini survey found that 82% of organizations plan to integrate AI agents within the next three years.

Amazon is also feeling the heat from its rivals. Anthropic made waves earlier this year with its direct screen-controlling Computer Use feature, OpenAI is pushing aggressively into task execution, and Google is reportedly readying agents that can book flights and buy products for users.

While Amazon rolled out conversational agents to Bedrock in July and updated Amazon Q Business with agent features just last week, those felt like incremental product tweaks. CEO Andy Jassy has promised a more capable, action-oriented Alexa, and this SF hub gives that vision a dedicated research engine.

Why it matters

We think this lab is Amazon admitting that prompt-and-response chatbots are a dead end for real enterprise productivity. The battle has moved to agency, and Amazon was lagging behind Anthropic and OpenAI in fundamental computer-use capabilities.

Pairing Adept's software navigation prowess with Pieter Abbeel's physical robotics pedigree is a clever hedge. While competitors are largely obsessing over browser tabs and spreadsheets, Amazon appears to be designing an operational nervous system that could eventually coordinate both digital workflows in AWS and automated machinery in its sprawling fulfillment centers.

That said, high hurdles remain. Regulators in both the US and Europe are actively probing these acqui-hire maneuvers for potential anticompetitive behavior. Furthermore, bridging the gap between an agent running experimental code in a sandbox and an autonomous system you trust with corporate credit cards and sensitive databases is a massive engineering leap. Amazon just assembled a dream team, but catching up to competitors who already have products in user hands will take serious hustle.