Build AI agents that use software like a human does.
Agents
Adept is the harder tool page to write honestly in 2026, because the company that gave us Adept is no longer really the company running it. Founded in 2022 by David Luan (ex-OpenAI, ex-Google Brain), Ashish Vaswani and Niki Parmar (both co-authors of the original "Attention Is All You Need" paper), and a small team of top-tier transformer researchers, Adept set out to build models that could use software the way a human does — see the screen, reason about the interface, and take actions like clicking, typing, and navigating. That research bet, called Action Transformer or ACT, produced ACT-1 in 2022 and ACT-2 in 2023, and for eighteen months Adept was one of the most watched agent-model labs in the industry.
Then in June 2024, Amazon licensed Adept's technology and hired most of the founding leadership, including CEO David Luan, into its AGI team. Adept-the-company continued to exist on paper — the corporate entity, a smaller staff, ongoing product commitments — but the center of gravity moved to Amazon. Since then Adept has operated more like a research lab-slash-legacy product than the aggressive frontier startup it was in its first two years. The website is still up at adept.ai; the product surface has narrowed; and the roadmap you'd have inferred in 2023 is not the roadmap that plays out.
We're including Adept in this catalog because the ideas it introduced — computer-using agents that see and act rather than just chat — are the ideas the whole industry has picked up. In 2026, Anthropic's Computer Use, OpenAI's Operator, Google's Project Mariner, and a growing pack of open-source alternatives are all building on the pattern Adept was first to name. If you're evaluating Adept in this environment, the honest question is: given the licensing deal and personnel move, is Adept still the right vendor for this problem, or has the frontier moved to the labs that hired its team?
We think the answer for most teams in 2026 is: Adept-the-product is worth knowing about historically, but for a production computer-using-agent decision today, you should evaluate Anthropic Claude Computer Use, OpenAI's Operator or Agent SDK, or an open-source alternative built on the same primitives. This page will explain why in more detail, but that's the takeaway if you're skimming.
ACT-Series Models. Adept's original bet: transformer models trained specifically to reason about user interfaces. ACT-1 and ACT-2 were demonstrated performing multi-step tasks in Salesforce, spreadsheets, and web applications by understanding screenshots and generating action sequences. Cutting-edge research when introduced.
Fuyu (Vision-Language Family). Adept released Fuyu-8B in 2023 as an open-source vision-language model — a simplified architecture designed for high-resolution image understanding. Fuyu was well-received in the research community and remains a reference model for compact multimodal transformers.
Persimmon (LLM). Adept's open-source large language model release, similarly research-driven. Not competitive at frontier scale but useful as an accessible base model for the community.
Adept Workflow Products. Product-side offerings built on ACT — enterprise-oriented offerings that let organizations automate repetitive software tasks (data entry, report pulling, cross-system operations). Adopted by a handful of design-partner enterprises rather than a broad customer base.
Enterprise Partnerships. Adept's go-to-market was always enterprise-first — not a self-serve SaaS but a design-partner motion with large organizations willing to fund custom deployments of computer-using agents.
Research Publications. The lab produced a stream of research artifacts — blog posts, model cards, benchmark contributions — that shaped the broader conversation about UI-native agents.
Post-2024 Reduced Surface. Since the Amazon licensing deal, the product roadmap has narrowed significantly. Whatever "Adept" ships in 2026 is more a maintained set of existing offerings than a growing product line. Prospective customers should verify current status directly with Adept before committing.
Adept's pricing has always been enterprise-quoted, not published. There is no self-serve tier, no monthly SaaS plan, and no public price list. In practice:
Design Partner / Enterprise (Custom) — Contracts historically negotiated per-deployment, per-workflow, or per-seat. Real contracts in Adept's active period clustered in the mid-five-figure to mid-six-figure annual range for enterprise deployments, depending on scope. Post-2024 this pricing is even harder to pin down without direct sales conversation.
Open-Source Model Releases (Free) — Fuyu and Persimmon were released under permissive licenses. You can download and use them from Hugging Face without paying Adept anything, at least for the versions that exist.
No Self-Serve — Unlike Anthropic Computer Use or OpenAI Operator, there is no "sign up, add a credit card, start using" path. Adept was never that shape of company.
| Feature | Open Models | Enterprise |
|---|---|---|
| Fuyu / Persimmon | Free (Apache-2.0) | Free |
| Managed agent product | N/A | Custom |
| ACT-based automation | N/A | Custom |
| Self-serve | No | No |
| Public price list | N/A | No |
| Starting price | $0 | Contact sales |
The realistic frame: if you're not a large enterprise with a specific computer-using-agent problem and a procurement team ready to write a custom contract, Adept was never for you. In 2026 that's more true than ever.
Pros.
Cons.
Reading about the history of computer-using agents. The archetype. If you want to understand how the industry got to Claude Computer Use and OpenAI Operator, Adept's blog posts and papers are the starting point. This is the honest primary use case in 2026 — as reference reading rather than as a vendor to buy from.
Downloading Fuyu or Persimmon for research work. The open-source releases are legitimately useful and still fine to build on. If you're doing multimodal research and want a compact vision-language base model, Fuyu-8B is a reasonable choice.
Existing enterprise customers with active contracts. If your organization already has an Adept deployment and it's producing value, staying the course through the contract term is defensible. Just plan the alternative in parallel; the Amazon-era Adept is unlikely to accelerate.
Anyone specifically evaluating pre-frontier UI-native transformer research. If you're a researcher looking at the ACT-1 / ACT-2 papers and want to build on the specific architectural approaches Adept explored, engaging with the historical research directly makes sense.
Bad fit — and this is the important one. Any team in 2026 evaluating a computer-using agent for production. The right evaluation set today is Anthropic Claude Computer Use, OpenAI Operator or the Agent SDK's computer-use variants, Google's emerging Project Mariner tooling, or open-source alternatives (browser-use, WebVoyager-style stacks) built on frontier vision-language models. Adept is not the sharp edge of the market anymore, and honesty requires saying so.
Anthropic Claude Computer Use. The current strongest general computer-using-agent capability. Claude sees the screen, reasons about the interface, generates click/type/scroll actions. Available via the Anthropic API, priced with Claude's usual per-token model. In 2026 this is the default recommendation for teams starting a computer-using-agent project.
OpenAI Operator / Agent Builder. OpenAI's browser-and-desktop-using agent products. Consumer-facing Operator, developer-facing agent SDK primitives that include computer-use variants. Aggressive roadmap, first-party integration with GPT-4o-class models.
Google Project Mariner. Google's browser-using agent, in extended preview as of early 2026. Deep integration with Chrome and Google's model stack.
Open-Source Computer-Use Agents. Projects like browser-use, WebVoyager, and SkyPilot-style stacks that combine a frontier vision-language model with a browser-automation harness. Slower and rougher than the frontier lab offerings but self-hostable and free.
AI Agent Frameworks. If your problem is agentic behavior more broadly rather than specifically computer-use, evaluate LangChain, CrewAI, AutoGPT, and Dify. Full comparison of the agent-framework landscape in our Dify vs LangChain vs CrewAI showdown.
Honorable mentions: Rabbit's Large Action Model (consumer device play, mostly a cautionary tale), MultiOn (browser-automation agent, active), and Cognition Labs' Devin (agentic coding, adjacent but different niche).
Given the reality of Adept's 2026 state, the "getting started" advice depends on why you're here.
If You Want to Read the Research. Start with Adept's blog at adept.ai — the ACT-1 and ACT-2 posts remain the clearest published descriptions of what computer-using agents were designed to do. Then read Anthropic's Computer Use documentation and OpenAI's agent product announcements to see how the pattern has evolved.
If You Want to Use Fuyu or Persimmon. Download from Hugging Face under Adept's org. Standard model card, standard inference setup. Treat as research tools, not production drop-ins.
If You Want to Buy Something From Adept. Contact sales via the adept.ai website. Expect a design-partner-style engagement, not a self-serve signup. Verify current staffing and roadmap directly; this is not a decision to make on reputation alone in 2026.
If You Want a Production Computer-Using Agent Today. Evaluate Anthropic Claude Computer Use first. Its per-token pricing is usually the cleanest starting point — a few dollars of API credits gets you a working prototype in an afternoon. If your task is browser-first, also evaluate the open-source browser-use library. If your workflow is Windows-heavy desktop automation, OpenAI's Operator and Agent SDK are the ones to test.
Practical tip: computer-using agents in 2026 are still not reliable enough to deploy autonomously on important workflows. Assume every deployment needs human-in-the-loop confirmation on any action that costs money, sends a message, or changes state you can't easily reverse. This is true across every vendor, not just one.
Is Adept still an independent company? The corporate entity still exists but in significantly reduced form after Amazon licensed the technology and hired most of the founding leadership in June 2024. The public-facing product surface is narrower than it was in 2023, and prospective customers should verify current status directly.
Are the Fuyu and Persimmon models still usable? Yes, the open-source releases remain available on Hugging Face under their original licenses. Community adoption of Fuyu is ongoing for research work.
Is Adept a good choice for a new computer-using-agent project? In our honest read, no — evaluate Anthropic Claude Computer Use, OpenAI Operator or Agent SDK, or open-source browser-automation stacks first. Adept was pathbreaking; the frontier moved.
Where did the technology go? Amazon licensed it and hired most of the team into its AGI initiative. Whether or when it re-surfaces as a shipped Amazon product is not publicly clear as of early 2026.
How much did Adept enterprise deployments cost? Not publicly disclosed. Real contracts in the company's active period clustered in the mid-five-figure to mid-six-figure annual range depending on scope. Current pricing requires direct sales contact.
Is Adept safe to build on? Depends on what "build on" means. Fuyu and Persimmon as open-source models are fine — you have the weights, they don't go away. An enterprise product contract requires more due diligence given the post-2024 uncertainty.
Who inherited Adept's role in the industry? Anthropic Claude Computer Use is the most direct heir in terms of technology and capability. OpenAI Operator, Google Project Mariner, and the open-source computer-use community are the broader field.
Adept was a genuinely important lab that named a pattern and produced early evidence that computer-using agents could work. The Fuyu and Persimmon open-source releases remain useful contributions to the research community. As a historical entity in the story of how AI agents got to where they are in 2026, Adept earned its place.
As a vendor to buy from in 2026, honesty compels the recommendation that most teams evaluate elsewhere first. The Amazon licensing deal and the hiring of the founding leadership moved the frontier of computer-using-agent work to Anthropic, OpenAI, Google, and the open-source community. Continuing to build on Adept's product surface is defensible for existing enterprise customers with active contracts, but a new project starting today should evaluate Claude Computer Use, OpenAI's Operator / Agent SDK, Google Project Mariner, or open-source browser-use stacks before committing to Adept.
If you're here to read the research, start with Adept's blog posts on ACT-1 and ACT-2 — they remain the clearest published descriptions of the ambition. If you're here to compare against a broader agent-framework decision (multi-agent orchestration, workflow builders, or code-first agent development), our Dify vs LangChain vs CrewAI comparison covers the general-purpose agent-framework landscape, and the Dify tool page covers the fastest path from concept to production for most non-computer-use agent problems.
Adept's story is a reminder that in AI in 2026, the frontier moves fast enough that a well-funded, well-staffed frontier lab can be redefined by a single acquisition. That's not a criticism of the research or the team — both were excellent. It's the honest context any potential customer deserves before writing a purchase order.