Adept

Adept builds large-scale machine learning models designed to perform actions in software rather than merely generating text. Its core research focus is the “action transformer,” a multimodal model that observes graphical user interfaces—buttons, forms, menus—and executes tasks by simulating keyboard and mouse inputs. The flagship model, ACT-2, is pitched as an enterprise agent that can navigate complex SaaS workflows, from updating CRM records to running spreadsheet formulas, without bespoke API integrations. Unlike traditional RPA tools that rely on scripted rules, Adept’s approach uses end-to-end neural networks trained to generalize across unseen software interfaces.

In addition to its agent stack, Adept released Fuyu, an open-source multimodal model capable of understanding images, charts, and documents alongside text. Fuyu was intended to demonstrate the lab’s computer-vision capabilities and to serve as a component in its broader agent architecture.

Following a 2024 deal in which Amazon hired the majority of Adept’s research and engineering staff and licensed its technology, the company now operates as a slimmed-down research lab. It retains some intellectual property and continues to pursue agent-action research, but its commercial trajectory has shifted from building a standalone product to managing licensed IP and a smaller research agenda.

Adept was founded in 2022 in San Francisco by David Luan, a former engineering leader at OpenAI and Google Brain, together with Ashish Vaswani and Niki Parmar, two of the co-authors of the seminal “Attention Is All You Need” Transformer paper. The startup emerged from stealth with an ambitious thesis: the next frontier after language generation was action—models that could directly manipulate software to complete knowledge-work tasks.

The company moved fast, releasing its first action model (ACT-1) and raising a massive Series B that valued it above $1 billion. However, co-founders Vaswani and Parmar departed in 2023 to pursue other ventures, and in mid-2024 Amazon effectively hollowed out the company by hiring the bulk of its technical team and licensing its models and agent technology.

Adept retained a smaller staff and some IP, but the exodus transformed it from a high-growth startup into a lean research holding operation with an uncertain path to market.

Adept once occupied a unique niche as the best-funded startup exclusively focused on action transformers, but the 2024 Amazon deal has effectively removed it as a standalone competitor to OpenAI, Anthropic, and Google. Those labs have since integrated computer-use and agentic capabilities directly into their frontier models—OpenAI’s Operator, Anthropic’s Computer Use, and Google’s Project Mariner—backed by far larger teams and distribution. Adept’s remaining advantage is its specialized IP around GUI-grounded action models and its relationship with Amazon, which may incorporate Adept’s research into AWS or Alexa workflows.

Against pure-play RPA incumbents like UiPath or Automation Anywhere, Adept’s neural approach was technically differentiated but never reached broad enterprise deployment. Today, Adept is less a competitor than a potential technology supplier or future acquisition target for larger platforms seeking proven agent IP.