A San Francisco startup with a modest online presence claims it has sold more than 33,000 hours of expert computer-use training data to some of the world's most advanced AI laboratories. Markov, part of Y Combinator's Summer 2026 batch, has positioned itself in a peculiar corner of the AI infrastructure market: capturing how humans actually use desktop software, then packaging those recordings for companies trying to teach machines to do the same.
The market for such data has grown quietly but rapidly. Markov's Y Combinator profile claims over 200,000 downloads on Hugging Face, while their website claims over 150,000, though exact figures vary depending on where you look. The discrepancy itself tells a story about the early, somewhat chaotic nature of this business. Both "33k+ hours sold" to frontier labs and "15,000+ hours sold" are listed on the Y Combinator page, indicating inconsistent reporting. Download counts shift from 150,000-plus on the company website to 200,000-plus on Y Combinator. Markov's founders declined to name their customers or clarify the inconsistencies.
What's certain is that AI agent developers face a stubborn problem. Current systems still fail most long-horizon computer tasks, the kind that require navigating multiple windows, tracking constraints, and recovering from unexpected interruptions. OSWorld 2.0, a benchmark released in June 2026, found the best agents complete only about one-fifth of workflows (20.6% task completion rate) under a 500-step limit. These are tasks that take humans an average of 1.6 hours and require roughly 318 tool calls with top models, according to a paper published on arXiv. The agents work, in other words, but they work slowly and unreliably.
That gap has created structural demand for what companies like Markov sell: verified traces of expert computer interactions. Gartner forecast spending on AI agent software to reach $206.5 billion in 2026 and $376.3 billion in 2027, according to a May 5, 2026 press release. But enterprises lack the supervised demonstration data needed to move agents from controlled pilots to production scale, analysts say. Someone has to capture how a skilled worker moves through Salesforce, or edits in Photoshop, or navigates AutoCAD. That someone, increasingly, is a startup with screen-recording software and a sales pitch to frontier labs.
Computer-use agents entered production during 2025 and 2026 at OpenAI, Anthropic, and Google. These are systems that control graphical user interfaces by interpreting screens, moving cursors, clicking buttons, and typing text. OpenAI's Computer-Using Agent powers a tool called Operator and became available in the Responses API in January 2025, scoring 38.1% on OSWorld and 58.1% on WebArena, the company wrote on its blog. Google announced built-in computer use in Gemini 3.5 Flash on June 24, 2026, for browser, mobile, and desktop environments. Anthropic published a system card in March 2026 noting Claude's ability to use computers "the way people do."
Despite those launches, deployment remains narrow. Gartner's inaugural Hype Cycle for Agentic AI, published in April 2026, noted only 17% of organizations have deployed agents, though more than 60% expect to within two years. Venture firm a16z wrote in an August analysis that computer-using agents are viable for production in "narrow, repeatable workflows" with stable business rules and missing APIs, but orchestration remains bespoke. Datasets, the firm added, matter.
The challenge lies in data provenance and quality. Frontier labs need synchronized screen recordings paired with mouse and keyboard events, narrations, task specifications, rubrics, and verified outputs to train and evaluate agents across desktop applications and browsers, OpenAI explained in January 2025 developer documentation. Datasets often include DOM and accessibility trees, bounding boxes, and dense screen parses for perception supervision. Standard Intelligence trained a general computer action model on an 11-million-hour video dataset, which a16z described as "an early signal that the step-by-step screenshot loop… is a solvable problem."
Regulatory pressure is tightening around data sourcing. The European Data Protection Board released guidelines on web scraping in the context of generative AI for public consultation in March 2026, signaling stricter expectations from Brussels. Ropes & Gray, a law firm, published guidance on May 27, 2026 detailing legal risks around scraping for AI training, including terms-of-service, intellectual-property, and privacy exposure. The OECD noted in an October 2025 report that Common Crawl and LAION aggregators face rising terms-of-service restrictions and robots.txt enforcement. The FTC has shown sensitivity to training on private recordings, referencing prior cases involving Alexa and Ring user data.
Against that backdrop, companies selling cleaned, rights-cleared expert demonstrations have multiplied. Datoric offers 250,000 real-world traces with commercial licensing. Paradigm Shift AI released a 3,100-task dataset and a recording tool called Captr. XLANG released OpenCUA and the AgentNet dataset. Nxtscape, General Data, Knavix, Andon Labs, and SuperAnnotate all advertise computer-use data pipelines targeting frontier labs. Turing published a case study describing the creation of more than 10,000 supervised GUI tasks to train a general-purpose computer-use agent.
Markov's open datasets show the shape of what frontier labs are buying. The company's largest public release, computer-use-large, contains 48,478 screen-recording videos spanning approximately 12,300 hours across AutoCAD, Blender, Excel, Photoshop, Salesforce, and Visual Studio Code, according to the dataset card on Hugging Face updated March 16, 2026. That dataset logged 16,439 downloads in the month before August 28, 2026, Hugging Face data shows. A second dataset, cad-1000-hours, provides roughly 1,022 recorded hours across 597 CAD, BIM, and analysis workflows, updated in August 2026 and drawing 56,917 downloads in the prior month. A third, gaming-500-hours, includes 776 workflows with synchronized clips and events, updated June 30, 2026.
The founders are Dev Mandal, who previously worked at Sarvam AI and studied at IIT Madras, and Harish Ashok, who founded Zenith, a robotics and AI tools effort, according to the company's Y Combinator listing. "Today, we're launching the world's largest open-source dataset of computer-use recordings. 10,000+ hours across Salesforce, Blender, Photoshop and more," Mandal wrote in a LinkedIn post around March 2026. In an earlier post, he wrote, "We're making a bet that advanced computer use agents which can navigate gui's and multiple software windows will be necessary to automate most economically valuable work."
OpenAI cited two named customers in January 2025 documentation. Unify uses agents to verify business footprints via online maps for property management. Luminai automated complex operational workflows for an enterprise with legacy systems, succeeding where robotic process automation had struggled for months, the company said. Automation Anywhere reported in a press release spanning May through August 2026 that it fulfilled more than one billion IT service requests via its autonomous desk and attributed 61% of Q4 bookings to AI. UiPath launched "UiPath for Coding Agents" in mid-2026, adding governance and integration for multi-vendor agents.
The gap between benchmark performance and human reliability remains wide, perhaps wider than the industry's public optimism suggests. OSWorld 2.0 tasks average 1.6 human hours and agents take 1.4 to 2.7 times more steps than necessary, the June 2026 paper found. Agents fail on constraint tracking, information that arrives mid-task, and hidden state recovery. The venture firm a16z wrote in August that "the frontier is shifting" as perception and action loops improve and video-trained action models emerge, but production deployments succeed only where workflows are high-volume, repetitive, with stable business rules and missing APIs.
Research released in 2025 and 2026 suggests synthetic and mined data may close part of the gap. VideoAgentTrek mines unlabeled screen-recorded videos at web scale, and WebSTAR introduces large-scale step-filtered synthetic supervision, according to papers published on arXiv. ScreenParse, released in 2026, offers 771,000 web screenshots with dense element annotations covering 21 million elements for complete screen parsing.
Cisco forecast in a May 2026 report that daily reliance on AI agents will become common between 2026 and 2035, with the majority of consumers using AI assistants by 2030. Gartner's Hype Cycle for Agentic AI noted market maturity remains uneven. Stratechery argued in a February 17, 2026 post that the agent's purpose is task accomplishment, not "using the computer for you" per se, reflecting UI and compute separation trends as assistants and agents mature.
OpenAI's system card for its Computer-Using Agent, updated through March 2025, emphasized that despite progress, performance on OSWorld remains far from human, requiring oversight. The company stressed dataset quality, safety mitigations, and environment isolation. Anthropic noted in its system card elevated susceptibility in GUI settings. The MIT 2025 Agent Index documented browser-agent vulnerabilities and the need for sandboxing and conduct policies. The UK Competition and Markets Authority and Department for Business and Trade published guidance on March 9, 2026 warning, "If an AI agent you use does something illegal, you are responsible."
Markov's reported sales and competitor datasets totaling millions of hours suggest frontier labs are stockpiling expert demonstrations before agents tackle longer, more complex desktop workflows at scale. Whether cleaned trajectories alone will close the human-agent gap remains an open question. The infrastructure market, for now, is betting on data quality as the next unlock. Contact for Markov is listed as [email protected].
