Adept

Between 2022 and 2024, Adept developed AI models and agents that operated software on a user's behalf. Its work progressed from ACT-1, a model that interacted with browser interfaces, to multimodal models and configurable enterprise workflows. In June 2024, the company shifted its focus toward agentic AI solutions while co-founders and some employees joined Amazon's AGI organization. Amazon licensed Adept's technology, models and datasets.[2][3][7][9]

In its June 28, 2024 update, Adept said that pursuing both foundation-model scaling and enterprise agent development would require substantial attention to fundraising. It chose to concentrate on agentic AI solutions using its existing models, agentic data, interaction software and infrastructure. Head of Engineering Zach Brock was announced as the incoming CEO, while Tim Weingarten remained Head of Product. Adept also reported production deployments with end-to-end workflows containing up to dozens of steps, without disclosing customer counts, revenue or a quantified success rate.[9]

The arrangement left Adept as a separate company. Semafor reported in August 2024 that ownership had not changed and approximately one-third of employees remained. Citing people briefed on the matter, it reported that investors would roughly recover the $414 million invested and Adept would receive around $25 million. These were reported prospective arrangements, rather than confirmation of completed payments. Semafor put Adept's most recent valuation at $1 billion.[12]

The models and interaction software retained by Adept grew out of its earlier effort to train neural networks to use software tools and APIs. That work began publicly with the company's April 2022 introduction.[2][9]

David Luan introduced Adept on April 26, 2022, identifying Ashish Vaswani and Niki Parmar as his co-founders. Vaswani and Parmar had contributed to the Transformer, while Luan brought experience at OpenAI and Google. The company's initial product vision was a user-supervised collaborator capable of operating existing software. At launch, Adept reported a nine-person team and $65 million raised, with Greylock, Addition and Root Ventures among the investors.[2]

Adept introduced ACT-1, its first large Action Transformer, in September 2022. A Chrome extension supplied a custom rendering of the browser viewport; the model acted through available interface elements by clicking, typing and scrolling, then observed the resulting state. In one Salesforce demonstration, a sentence replaced a task Adept said would ordinarily require ten or more clicks. Other demonstrations involved spreadsheets, online research and combining tools. The videos were sped up, so their playback did not establish actual task latency.[3]

On March 14, 2023, Adept announced $350 million raised in a Series B led by General Catalyst and co-led by Spark Capital. It said the financing would support initial product launches, model training and hiring. Together with the launch announcement, this amounted to $415 million in announced rounds; Semafor's later report used the slightly different figure of $414 million invested.[4][2][12]

In September 2023, Adept released Persimmon-8B, an approximately eight-billion-parameter language model trained from scratch with a 16K context. Its code and weights used an Apache license. The release was a raw model without additional fine-tuning or output-control postprocessing, with unused embeddings reserved for multimodal extensions. Adept presented it as an early result of the model-scaling work behind its agent products.[5]

October's Fuyu-8B release introduced a multimodal architecture built around a decoder-only Transformer without a separate image encoder. Image patches were linearly projected into the first Transformer layer, with raster ordering and image-newline tokens accommodating different image resolutions. The released checkpoint used a CC-BY-NC license, unlike Persimmon's Apache license. Adept separately previewed internal Fuyu-based capabilities for OCR, interface localization and interface question answering; those previews did not establish that every capability was present in the public checkpoint.[6]

In January 2024, Adept announced Fuyu-Heavy, a scaled multimodal model it planned to use in its enterprise product. Adept reported scores of 48.3 on MMMU, 76.2 on VQAv2, 81.2 on AI2D and 75.4 on ChartQA. These measured benchmark performance, rather than agent workflow success. The company described architecture and training changes to address instability and credited Microsoft Azure for its training cluster and service, and NVIDIA for model-efficiency and networking collaboration.[8]

Training also produced a concrete infrastructure investigation. In a September 2023 account, Erich Elsen, Curtis Hawthorne and Arushi Somani traced abrupt NaNs and subtler loss changes to silent hardware errors. Deterministic reruns returned different results despite normal ECC counters. Testing individual nodes isolated a problematic machine; replacing it enabled a run lasting weeks without a NaN. Parameter and optimizer checksums revealed correlated corruption, and the team explained how an all-gather operation could spread a single error across downstream machines.[11]

Adept began rolling out Adept Experiments in November 2023. Its first experiment, Workflows, let users configure web tasks in natural language. It used ACT-2, fine-tuned from the Fuyu family for interface understanding, knowledge-worker data comprehension and action taking. Workflows observed pixels and acted through coordinates and keystrokes. Users could inspect clicking and typing steps, supply missing information, and choose visible stepwise execution or automatic execution of the sequence. Examples included creating Shopify discount codes and manipulating a canvas-based maps interface. Adept reported its first enterprise customers at the time.[7]

Adept continued explaining its product after the Amazon transition. In August 2024, it described Adept Workflow Language (AWL), a proprietary syntactic subset of JavaScript ES6 whose functions invoke model calls. A prescriptive click instruction could be translated into bounding-box coordinates for web actuation. The natural-language act() function instead invoked a reasoning loop that planned, observed results and recursively issued commands. Combining these approaches let workflow authors specify some steps precisely while giving the agent discretion over others.[10]

One AWL example extracted an attendee's information from a PDF and created a HubSpot lead. Adept adapted it to create a Salesforce lead by changing the destination URL, salutation and application terminology. The example illustrated how the same workflow could be adjusted across business applications; it was not a measured customer deployment outcome.[10]