Prentis, Backed by Reid Hoffman and Marc Pincus, Seeks $100M to Build AI Agents That Control Computers

When AI Stops Chatting and Starts Clicking

For two years, the dominant metaphor for AI has been the conversation. You type, it responds. Reid Hoffman, Mark Pincus, and a 31-year-old serial entrepreneur named Ritankar Das are betting that the next phase drops the conversation entirely. Their startup, Prentis, is building agents that take over the keyboard and mouse.

Prentis, launched in April and described in its materials as an AI research lab focused on computer use models, is in talks to raise $100 million at a $1 billion valuation, according to two people familiar with the discussions 1. The funding is not the story. The story is what the money builds: agents that learn how office workers navigate documents and systems, then replicate those workflows by directly controlling the computer 1.

This is a different category of risk from a chatbot that fabricates a source. A conversational AI gives you a wrong answer. An agent that controls your operating system takes a wrong action. When it clicks the wrong button in a production database, files the wrong customs declaration, or sends the wrong insurance adjustment, the damage is operational, not informational.

Prentis says it has signed contracts worth up to $50 million with customers including a healthcare management service organization and several manufacturers, according to the same two sources 1. Investor materials obtained by TechCrunch project $75 million in annualized run rate by the third quarter of this year. But Prentis's own pitch deck qualifies those figures as "estimated annualized value based on a contracted fee equal to 20% of savings realized, not recognized revenue," adding that they are "performance-dependent and subject to final execution" 1. That is a projection of a projection, contingent on customers actually saving money and Prentis capturing a fifth of it.

The company claims its Hive-32B model outperforms OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on two computer-use benchmarks: WindowsAgentArena, which measures task completion on real Windows applications, and ScreenSpot-v2, which tests a model's ability to locate on-screen controls 1. Prentis also says its model costs roughly 10 times less per task than frontier APIs 1. TechCrunch has not independently verified the benchmark results 1.

The competitive field is already dense. Anthropic bought the Seattle computer-use startup Vercept earlier this year, folding in its founders and shutting down its product 1. OpenAI is building its own agents. Mira Murati's Thinking Machines Lab is pursuing the same category 1. Each competitor has resources that make a $100 million raise look like pocket change.

Das brings credentials that justify the attention. He graduated from UC Berkeley at 18 with a double major in bioengineering and chemical biology, was the university's youngest University Medalist in more than a century, earned a master's at Oxford, and dropped out of an AI PhD program at Cambridge 1. He previously founded Titan, a holding company that has launched and operated AI companies including Tala Health, a virtual care provider that raised a $100 million seed round 1. Prentis has already hired more than 25 employees, including researchers from OpenAI, Google DeepMind, Meta, Tencent, and Alibaba, according to its website 1.

Hoffman and Pincus are background operators here. Hoffman, the LinkedIn co-founder and Greylock partner, recently left Microsoft's board to focus on Manas AI, a drug-discovery startup 1. Pincus, the Zynga founder, runs Reinvent Capital, where Hoffman serves as a senior adviser 1. For both, Prentis is a bet alongside other bets.

The unknown is not whether Prentis can hire talent or close deals. It is what happens to the desktop software ecosystem when the primary user of an application is an agent, not a human. Software vendors design for human interaction: confirmation dialogs, undo buttons, rate limits, session timeouts. Agents interact differently. They click faster. They navigate non-linearly. They may trip bot-detection safeguards or bypass them in ways that open new vulnerabilities. The application stack was not architected for a machine actor operating at machine speed, and adapting it will require rethinking security models that took decades to build.

Hoffman and Pincus are making a specific wager: the next platform layer is not a new application but a replacement for the person at the keyboard. If they are right, every company that builds desktop software faces a choice: design for the agent or lose the employer who deploys one. If they are wrong, that $1 billion valuation rests on contracts measured in projected savings rather than recognized revenue.