Adept AI
Last reviewed
Apr 26, 2026
Sources
35 citations
Review status
Source-backed
Revision
v4 · 3,580 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 26, 2026
Sources
35 citations
Review status
Source-backed
Revision
v4 · 3,580 words
Add missing citations, update stale details, or suggest a clearer explanation.
Adept AI (also known as Adept AI Labs) is an American artificial intelligence company founded on January 5, 2022, in San Francisco, California. The company was co-founded by David Luan (CEO, former VP of Engineering at OpenAI), Ashish Vaswani, and Niki Parmar, with Vaswani and Parmar being two of the eight co-authors of the landmark 2017 paper "Attention Is All You Need," which introduced the transformer architecture that underpins virtually all modern large language models. Additional co-founders included Augustus Odena, Maxwell Nye, Erich Elsen, Anmol Gulati, Fred Bertsch, and Kelsey Szot [1][2].
Adept set out to build AI systems that could take actions in the digital world, operating software tools the way a human would. Rather than focusing on chatbots or text generation, the company pursued what it called "action models": AI agents trained to interact with any software application through clicks, keystrokes, and scrolling. This vision attracted $415 million in total funding across two rounds and valued the company at over $1 billion at its peak [3][4].
However, Adept's trajectory shifted dramatically in June 2024, when Amazon executed an acqui-hire deal that brought CEO David Luan and a majority of Adept's staff into Amazon's artificial general intelligence (AGI) division. The deal drew scrutiny from the U.S. Federal Trade Commission (FTC) as part of a broader investigation into whether large technology companies were structuring talent acquisitions to circumvent merger notification requirements [5][6].
Adept AI emerged from stealth on April 26, 2022, announcing $65 million in Series A funding led by Greylock and Addition. The founding team brought together researchers from some of the most prominent AI organizations in the world. David Luan had served as VP of Engineering at OpenAI from 2017 to 2020, where he helped build and ship GPT-2, GPT-3, CLIP, and DALL-E. After leaving OpenAI, Luan spent roughly 15 months as a Director at Google Brain, leading large-scale model training efforts. Ashish Vaswani and Niki Parmar were both former Google Brain researchers who had been instrumental in developing the transformer architecture [7][8].
The company's founding thesis was that the most impactful near-term application of AI would not be generating text or images but rather teaching AI to use existing software tools. Adept described this as building "general intelligence" in the form of a system that could do anything a human can do in front of a computer. The Series A round attracted notable angel investors including Andrej Karpathy (then head of Tesla Autopilot), Dara Khosrowshahi (CEO of Uber), Scott Belsky (founder of Behance), and Howie Liu (founder of Airtable) [7].
On September 14, 2022, Adept publicly demonstrated its first model, the Action Transformer (ACT-1). This was a large-scale transformer model trained to use digital tools, connected to a Chrome browser extension that allowed it to observe what was happening on screen and take actions such as clicking, typing, and scrolling [9].
The ACT-1 demo showed the model performing a range of tasks that required understanding user intent and interacting with web applications. In one demonstration, a user typed a natural language instruction into a command bar, and ACT-1 navigated to a real estate website, entered search criteria, applied filters, and browsed listings. In another example, ACT-1 worked within spreadsheet software to manipulate data based on high-level instructions. The model demonstrated the ability to compose multiple tools together, look up information it did not know, and correct its own mistakes after a single piece of human feedback [9][10].
ACT-1 attracted significant attention in the AI community because it represented a different direction from the chatbot-focused approaches of most AI startups at the time. Rather than generating text responses, ACT-1 was designed to take real actions in real software environments. Adept described it as a "foundation model for actions" that was trained to use every software tool, API, and web application that exists [9].
In 2023, co-founders Ashish Vaswani and Niki Parmar departed Adept to launch a new startup called Essential AI. Essential AI focused on building enterprise software powered by large language models and went on to raise $64.5 million from investors including Thrive Capital, Google, Franklin Templeton Investments, and NVIDIA. The departure of two of Adept's most high-profile co-founders, both credited as inventors of the transformer architecture, was a notable setback for the company. However, Adept continued to operate under Luan's leadership with the remaining co-founders and engineering team [11][12].
On March 14, 2023, Adept announced a $350 million Series B funding round co-led by General Catalyst and Spark Capital. The round valued the company at approximately $1 billion, granting it unicorn status. Additional investors included Addition, Greylock, Atlassian Ventures, Microsoft, NVIDIA, Workday Ventures, Frontiers Capital, PSP Growth, SV Angel, A.Capital, and Caterina Fake. This brought Adept's total funding to approximately $415 million [3][4][13].
The Series B reflected enormous investor enthusiasm for AI agent technology at a time when ChatGPT had recently launched and venture capital was flooding into AI startups. Adept's pitch was differentiated: rather than building another chatbot or language model for text generation, the company was training models to automate knowledge work by directly operating enterprise software [13].
On September 7, 2023, Adept open-sourced Persimmon-8B, a text-only language model with approximately 9.3 billion parameters. At the time of its release, Adept described it as the most powerful fully permissively-licensed language model under 10 billion parameters. Key features of Persimmon-8B included [14][15]:
The model used a standard decoder-only transformer architecture with several modifications, including the squared ReLU activation function, rotary positional encodings, and layer normalization applied to the Q and K embeddings before the attention calculation [14].
On October 17, 2023, Adept released Fuyu-8B, an open-source multimodal model designed from the ground up for digital agents. Unlike most multimodal models at the time, Fuyu-8B used a simplified architecture: a vanilla decoder-only transformer with no specialized image encoder. Image patches were linearly projected directly into the first layer of the transformer, bypassing the embedding lookup entirely [16][17].
This architectural simplicity gave Fuyu-8B several practical advantages. It could support arbitrary image resolutions without preprocessing or resizing. It could process large images and return responses in under 100 milliseconds. And it was straightforward to scale, train, and deploy compared to architectures that relied on separate vision encoders [16].
Fuyu-8B was released under a CC-BY-NC (Creative Commons Attribution-NonCommercial) license and was made available on Hugging Face. The model demonstrated strong performance on visual question-answering and natural image captioning benchmarks, and it was specifically optimized for understanding user interfaces, answering questions about graphs and diagrams, and performing fine-grained localization on screen images [16][17].
In late January 2024, Adept announced Fuyu-Heavy, a larger multimodal model in the Fuyu family. The company described Fuyu-Heavy as the world's third-most-capable multimodal model at the time, behind only GPT-4V and Gemini Ultra, which were estimated to be 10 to 20 times larger. On the MMMU (Massive Multi-discipline Multimodal Understanding) benchmark, Fuyu-Heavy scored higher than Gemini Pro. On standard text-only evaluations, it performed roughly on par with Gemini Pro and outperformed it on the MMLU benchmark. In chat evaluation, it performed comparably to Claude 2.0 [18][19].
Fuyu-Heavy was not open-sourced. Instead, it served as the foundation for Adept's internal product development and was later used in the ACT-2 model that powered Adept's enterprise workflow automation tools [18].
Adept launched a product initiative called "Adept Experiments," with Workflows as its first offering. Workflows was powered by ACT-2, a multimodal model fine-tuned from the Fuyu family and optimized for UI understanding, knowledge worker data comprehension, and action execution. The system worked as an end-to-end multimodal AI agent that perceived the screen directly via pixels and acted on the computer through coordinates and keystrokes [20][21].
Adept began working with enterprise customers, demonstrating the ability to learn new workflows in minutes and execute complex multi-step tasks across multiple software applications. The company reported over 95% reliability for workflows developed in direct collaboration with enterprise clients. Use cases included hiring process automation, document extraction, database updates, and multi-application data workflows [20][21].
On June 28, 2024, Amazon announced that it had hired David Luan and several other Adept co-founders, including Augustus Odena, Maxwell Nye, Erich Elsen, and Kelsey Szot, to join its AGI team led by Rohit Prasad. As part of the arrangement, Amazon obtained a non-exclusive license to Adept's technology, including its family of multimodal models and certain datasets [5][22][23].
The deal was not structured as a traditional acquisition. Amazon stated that it did not consider the arrangement an acquisition because it was not interested in owning Adept's business and technology outright. Instead, approximately two-thirds of Adept's roughly 100 employees joined Amazon, while about 20 employees remained at Adept. Amazon paid approximately $25 million in licensing fees [5][6][24].
The deal had significant implications for Adept's investors. Firms including Greylock and General Catalyst had collectively invested $414 million in the company. According to reporting by Semafor, the co-founders arranged for investors to roughly recoup their investment using funds paid by Amazon. However, investors received approximately their money back rather than the substantial returns they had expected from a billion-dollar AI company. For venture capital firms that had entered at the Series B valuation, getting their money back represented a significant disappointment compared to the typical expectations for a unicorn investment [24].
Adept retained about one-third of its workforce and approximately $25 million in operating capital. Zach Brock, previously Adept's Head of Engineering, took over as CEO. Tim Weingarten remained as Head of Product. The reconstituted company shifted its focus to enabling agentic AI solutions, working with partners to bring agentic capabilities to their products rather than spending resources on training expensive foundation models [25][26].
The Amazon-Adept deal attracted scrutiny from the U.S. Federal Trade Commission (FTC), which launched an informal inquiry into the arrangement in July 2024. The FTC sent a list of questions to Amazon seeking details about its relationship with Adept. Regulators were examining whether the mass hiring effectively amounted to an acquisition of Adept that circumvented merger notification rules under the Hart-Scott-Rodino (HSR) Act [6][27].
The investigation was part of a broader FTC effort to scrutinize so-called "reverse acqui-hires" in the AI sector. Similar deals were also under review, including Microsoft's March 2024 arrangement with Inflection AI, in which Microsoft paid $650 million to license Inflection's technology and hired most of its 70-person workforce, including co-founder Mustafa Suleyman [6][28].
In January 2025, the FTC published a staff report on corporate partnerships between large cloud service providers (Alphabet, Amazon, and Microsoft) and AI developers, as part of a broader Section 6(b) study. The report explicitly referenced the Amazon-Adept and Microsoft-Inflection arrangements as examples of deals that raised competition concerns. The report noted that such partnerships may impact access to computing resources and engineering talent, increase switching costs, and provide cloud service provider partners with access to sensitive technical and business information [28].
U.S. Senator Ron Wyden (D-Oregon) also cited Amazon's deal with Adept in a letter to the FTC and Department of Justice, arguing that technology companies were licensing technology or making acqui-hires specifically to avoid antitrust scrutiny. As of March 2026, no formal enforcement action had been publicly announced in connection with the Amazon-Adept deal [27].
In December 2024, Amazon formally established the AGI Lab, a small research group based in San Francisco focused on "long-term research bets" in AI agent technology. David Luan was named to lead the lab as Vice President of Autonomy. Pieter Abbeel, a robotics researcher who had joined Amazon through a similar licensing-and-hiring deal with Covariant, worked closely with Luan's team [29][30].
The lab's first major release was Amazon Nova Act, an AI model and developer toolkit for building agents that can perform tasks autonomously in web browsers. Amazon unveiled Nova Act in preview form on March 31, 2025. The model was engineered to handle interface elements that commonly cause problems for AI agents, such as dropdown menus, date pickers, and popup dialogs. Amazon reported that Nova Act achieved over 90% accuracy on internal evaluations of these challenging UI interactions and reached best-in-class performance on benchmarks including ScreenSpot and GroundUI Web [31][32].
Nova Act drew a clear line from Adept's original vision of action models to Amazon's product ambitions. The technology was adopted by customers including Hertz, 1Password, and Amazon.com itself [31].
In February 2026, David Luan announced his departure from Amazon, less than two years after joining through the Adept acqui-hire. In a LinkedIn post, Luan said he was leaving "to cook up something new," adding that "with AGI so close, I decided to spend 100% of my time on teaching AI systems brand new capabilities" [33][34].
Luan's departure followed a major reorganization of Amazon's AGI division in late 2025, which placed the unit under Peter DeSantis, a 27-year Amazon veteran and Senior Vice President in Amazon Web Services. By the time of Luan's departure, four of the five Adept co-founders who had joined Amazon had already left the company [33][34].
| Model | Type | Parameters | Release Date | License | Key Features |
|---|---|---|---|---|---|
| ACT-1 | Action model | Undisclosed | September 2022 | Proprietary | First action transformer; operated via Chrome extension; clicked, typed, scrolled in real software |
| Persimmon-8B | Text LLM | ~9.3B | September 2023 | Apache 2.0 | 16K context window; competitive with LLaMA 2 on 37% of the training data |
| Fuyu-8B | Multimodal | ~8B | October 2023 | CC-BY-NC | No image encoder; arbitrary image resolution; sub-100ms latency for large images |
| Fuyu-Heavy | Multimodal | Undisclosed | January 2024 | Proprietary | Third-most-capable multimodal model at launch; outperformed Gemini Pro on MMMU |
| ACT-2 | Action model | Undisclosed | 2024 | Proprietary | Fine-tuned from Fuyu family; powered Adept Workflows product; optimized for UI understanding |
| Round | Date | Amount | Lead Investors | Valuation | Key Details |
|---|---|---|---|---|---|
| Series A | April 2022 | $65 million | Greylock, Addition | Undisclosed | Came out of stealth; notable angels included Andrej Karpathy, Dara Khosrowshahi |
| Series B | March 2023 | $350 million | General Catalyst, Spark Capital | ~$1 billion | Achieved unicorn status; Microsoft and NVIDIA among investors |
| Total | ~$415 million |
| Name | Role at Adept | Background | Post-Adept |
|---|---|---|---|
| David Luan | Co-founder, CEO | VP of Engineering at OpenAI (2017-2020); Director at Google Brain | Led Amazon AGI Lab (2024-2026); departed February 2026 |
| Ashish Vaswani | Co-founder, Chief Scientist | Co-author of "Attention Is All You Need"; Google Brain | Left Adept in 2023; co-founded Essential AI |
| Niki Parmar | Co-founder, CTO | Co-author of "Attention Is All You Need"; Google Brain | Left Adept in 2023; co-founded Essential AI |
| Augustus Odena | Co-founder | Google Brain researcher | Joined Amazon (June 2024) |
| Maxwell Nye | Co-founder | AI researcher | Joined Amazon (June 2024) |
| Erich Elsen | Co-founder | AI researcher | Joined Amazon (June 2024) |
| Kelsey Szot | Co-founder | AI researcher | Joined Amazon (June 2024) |
| Zach Brock | CEO (post-Amazon deal) | Head of Engineering at Adept | Took over as CEO after co-founders departed |
Adept AI occupies a distinctive place in the history of the AI industry for several reasons.
First, Adept was one of the earliest companies to articulate and pursue the vision of AI agents that interact with software through the same interfaces humans use. While ChatGPT and other chatbots captured public attention, Adept was building models that could click buttons, fill forms, and navigate applications. This "action model" approach anticipated the broader industry shift toward AI agents that began in earnest in 2024 and 2025, when companies like OpenAI, Google, Anthropic, and Amazon all invested heavily in agentic AI capabilities.
Second, Adept's open-source contributions, particularly Persimmon-8B and Fuyu-8B, advanced the state of multimodal AI research. Fuyu-8B's simplified architecture, which eliminated the need for a separate image encoder, influenced subsequent work on vision-language models and demonstrated that competitive multimodal performance could be achieved with more streamlined designs.
Third, the Amazon acqui-hire became a landmark case in the debate over how large technology companies acquire AI talent. Alongside Microsoft's deal with Inflection AI, the Adept deal prompted regulatory scrutiny and legislative attention. The FTC's investigation and subsequent Section 6(b) report helped establish a framework for evaluating whether talent acquisitions and licensing deals are effectively acquisitions in disguise. These cases influenced how regulators approach similar transactions across the AI industry.
Finally, Adept's story illustrates the challenges facing well-funded AI startups competing against technology giants with vastly larger resources. Despite raising $415 million, Adept ultimately found that the costs of training frontier AI models and competing for talent against companies like Amazon, Google, and Microsoft were prohibitive. The acqui-hire pattern that consumed Adept was not unique; it reflected a broader dynamic in which several promising AI startups were effectively absorbed by incumbent technology companies during 2023 and 2024 [35].
As of early 2026, Adept continues to operate as an independent company under CEO Zach Brock with approximately 20 employees. The company has shifted its focus from training foundation models to building solutions that enable agentic AI, leveraging its existing models, agentic data, web interaction software, and custom infrastructure. Adept works with partners to bring agentic capabilities to their products rather than competing directly as a model provider [25][26].
Amazon continues to develop the AI agent technology that originated at Adept through the Nova Act product line and its broader AGI division [31].