Adept

What is Adept?

Adept is a San Francisco AI company that built multimodal agent models able to read a screen and act inside web software on a user’s behalf. Its aim was to let people describe a task in plain language and have the agent click, type and scroll through the tools needed to finish it.

The company was founded in 2022 by David Luan and announced ACT-1, a transformer trained to use a web browser, in September of that year. A $350 million Series B followed in March 2023. Adept then released two open models, Persimmon-8B and Fuyu-8B, and in January 2024 introduced Fuyu-Heavy, a larger multimodal model built for digital agents.

In June 2024 Luan and several co-founders joined Amazon’s AGI organisation, and Amazon licensed Adept’s agent technology, multimodal models and datasets. Zach Brock, previously head of engineering, became chief executive, and the remaining team said it would concentrate on agent products for businesses. The official site still shows only blog posts, the newest dated August 2024. A January 2026 business directory describes a leaner, product-focused Adept, but we could not confirm a live product on adept.ai.

Key Features

Adept’s most important features are its own agent models and its workflow tooling, because together they were designed to operate real business software through the interface rather than through APIs.

• Screen-reading agent models: Adept’s models look at a rendered page and choose UI elements to act on, step by step. ACT-1 was shown condensing a Salesforce task that normally needs more than ten clicks into one typed instruction.
• Adept Workflow Language: AWL is a custom language for composing multimodal web interactions into repeatable tasks. Teams describe a multi-step process once, and the agent then follows those instructions across the pages and forms involved.
• Fuyu-Heavy: This January 2024 model was built for interface understanding and reasoning over charts and documents. Adept reported 48.3 on the MMMU benchmark and 81.2 on AI2D, and said it would power the enterprise product.
• Fuyu-8B open model: Fuyu-8B is a decoder-only model that feeds image patches straight into the transformer without a separate image encoder. It accepts images at any resolution and, by Adept’s account, answers on large images in under 100 milliseconds.
• Persimmon-8B: Released in September 2023, this is a permissively licensed language model for developers who want a small, open base model. It gave researchers a way to inspect Adept’s work without needing commercial access.
• Adept Experiments: Launched in November 2023, this was a web-based workflow builder where users set up tasks in natural language. It ran on ACT-2, a Fuyu-derived model, and was opened to users on a rolling basis.

Pros

The main reason to look at Adept is its research record in agents that use software, which Amazon considered valuable enough to license in 2024.

• Early agent pioneer: Adept demonstrated a browser-driving agent in 2022, well before such features became common in mainstream assistants. Buyers comparing agent vendors gain a team with years of focused work on this problem.
• Open research models: Fuyu-8B and Persimmon-8B remain available to download and test at no cost. Researchers and developers can study screen and chart understanding without signing a contract or sharing data.
• Interface-level automation: Agents act through the visible interface, so they can work with legacy tools that lack modern APIs. This suits organisations running older CRM, ERP or supplier portals that are hard to integrate.
• Security groundwork: Adept announced SOC 2 Type 1 compliance in May 2024. Enterprise buyers therefore have at least a formal report on the design of its security controls to request.
• Long workflows in beta: Adept’s June 2024 update said beta deployments automated end-to-end workflows of up to dozens of steps. That suits rule-heavy back-office processes rather than single one-off actions.

Cons

Adept’s biggest drawback is uncertainty about its current status, since the founders left for Amazon in 2024 and the official site has published nothing since.

• Unclear product status: The newest post on adept.ai dates from August 2024, and several older blog links now return errors. You should confirm that a supported product exists before planning any project around it.
• No self-serve access: There is no sign-up form, free plan, trial or app download on the official site. Individuals and small teams cannot try the commercial product before speaking to the company.
• Hidden pricing: Adept has never published prices for its enterprise agents. Budgeting is difficult, and comparisons with priced alternatives such as UiPath or OpenAI’s agent tools need a sales conversation first.
• Talent loss: Most of the founding team moved to Amazon, and Amazon holds a licence to the core technology. That raises fair questions about how fast the remaining company can ship and support new features.
• Non-commercial open weights: Fuyu-8B is released under a CC-BY-NC 4.0 licence, which rules out commercial use. Businesses wanting a free model for production work will need to look elsewhere.

Who is Using Adept?

Adept was built for large organisations with repetitive software work, while researchers mainly use its open models.

• Large enterprises: Operations leaders have explored Adept to automate repetitive work in legacy web tools. Directories cite finance, healthcare and supply chain processes as typical targets.
• Back-office teams: Staff handling data entry, invoices or document extraction were the intended users of Adept Workflows. The agent copies information between systems that do not talk to each other.
• AI researchers: Academics study Adept’s papers and models to understand multimodal agents. Fuyu-8B is a common reference point for simple image-and-text architectures.
• Developers: Engineers test Fuyu-8B on screenshots, charts and OCR-style questions. It helps them prototype interface-reading features before choosing a commercial model.

Pricing

Adept publishes no price list, and any commercial use appears to require contacting the company directly (checked on adept.ai, October 2026).

Futurepedia and Toolify list Adept as contact for pricing but do not mention the 2024 leadership change.

Disclaimer: We could not confirm a live commercial offer, so check terms and availability with Adept before planning a purchase.

What Makes Adept Unique?

Adept trained its own models to operate software through the visual interface rather than wrapping a general chatbot around existing APIs. ACT-1 read a custom rendering of the browser viewport and picked the elements to act on, an approach that predates most current computer-use agents by about two years.

Its Fuyu design is also unusual. Instead of bolting an image encoder onto a language model, Fuyu projects image patches directly into the transformer, which lets it handle screenshots of any size. Amazon’s decision to license this stack in 2024 is the clearest outside signal of how the work was valued.

Compatibilities and Integrations

Adept works through the browser interface, so compatibility depends on the web software you want to automate rather than on formal connectors.

• Platform: Agents act inside a web browser; ACT-1 was demonstrated through a Chrome extension.
• Business software: Demos covered Salesforce and spreadsheets, and directories report use with CRM and ERP tools.
• Developer access: Fuyu-8B weights are published on Hugging Face; we found no public API for the enterprise product.
• Security: A SOC 2 Type 1 report was announced for enterprise buyers in May 2024.

Adept Tutorials

Without a self-serve app, your route into Adept depends on whether you want a business product or research access.

1. Read the June 2024 update on adept.ai to understand the current company structure.
2. Contact Adept to ask whether enterprise agent work continues and what support is offered.
3. Request a current security report before sharing any customer data.
4. For research, download Fuyu-8B from Hugging Face and check its non-commercial licence.
5. Test the model on your own screenshots, charts and documents.
6. Compare results with actively maintained agents such as Amazon Nova Act before committing.

How We Rated It

Adept lands at 3.3 because capable agent research is offset by an unclear status, no self-serve access and hidden pricing.

Summary

Adept suits researchers and large enterprises interested in agents that work directly through software interfaces. Its biggest limitation is uncertainty, because the founders left in 2024 and the official site has announced nothing new since. We would not plan a project around it without written confirmation of support and pricing. Its open models still have research value, provided you respect the non-commercial licence. If you only want to experiment, start with the free Fuyu-8B weights.