The Next Battlefront for AI Is Not Code — It’s Your Desktop
If you’ve been following AI news the way I have — obsessively, with a mix of excitement and mild existential dread — you’ve noticed something shifting lately. It’s not about who has the bigger model anymore. It’s about what these models can actually do with a computer.
Two stories broke this week that, taken together, tell us where the industry is really heading. First, Prentis, a new AI research lab co-founded by Reid Hoffman and Mark Pincus, is in talks to raise $100 million at a $1 billion valuation — not to build a better chatbot, but to build AI agents that can navigate office workflows the way a human would. Second, Cognition acquired Poke (the AI assistant you text like a friend) in a “low nine figures” deal, specifically to bring personality and conversational style to its coding agent Devin.
These aren’t random funding rounds. They’re signals that AI is entering a new phase — one where the ability to use software and the ability to be likeable matter as much as raw intelligence.
What Is Prentis Building?
Prentis launched in April 2026 with a clear thesis: the biggest AI opportunity isn’t generating code — it’s automating the routine computer tasks that office workers do every day. Processing insurance claims. Handling customs duty refund exceptions. Chasing down paperwork across spreadsheets, email threads, and legacy systems.
The company’s Hive-32B model is purpose-built for what they call “computer use” — navigating real Windows applications, clicking buttons, filling forms, reading documents. And according to their benchmarks, it outperforms OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6 on two industry tests: WindowsAgentArena (end-to-end task completion in real Windows apps) and ScreenSpot-v2 (locating the right on-screen control).
Here’s the part that caught my attention: Prentis claims its model costs roughly 10 times less per task than frontier APIs. That’s the kind of economics that makes enterprises pay attention. They’ve already signed contracts worth up to $50 million with customers in healthcare, manufacturing, and retail.
The team is stacked — over 25 employees, including researchers from OpenAI, Google DeepMind, Meta, Tencent, and Alibaba. The CEO, Ritankar Das (31 years old, UC Berkeley’s youngest University Medalist in a century, Gates Cambridge Scholar at Oxford), is the same founder who built Titan, a holding company that’s launched several AI-powered healthcare startups. These are not amateurs.
Why Personality Matters: Cognition’s Poke Acquisition
On the same day the Prentis story broke, Cognition announced it had acquired Poke — the AI assistant known for texting you like a friend — for a “low nine figures.” On the surface, this looks like two completely different stories. One is about serious enterprise automation. The other is about a chatbot with personality.
But they’re connected by a deeper thread: how AI interacts with us is becoming as important as what it can do.
Poke is different from most AI assistants. Instead of responding like a search engine results page, it chats like a friend — slang, humor, personality. Users exchanged over 100 million messages on the platform in just three months. It became the first AI agent approved to run on Apple’s Messages for Business platform.
Cognition’s co-founder Scott Wu put it well: “The Interaction team has built an agent that people love: it’s proactive, it knows you, and it’s fun to talk to. That’s exactly how working with Devin should feel.”
The long-term vision? Poke orchestrates multiple Devin coding sessions (right now Devin can only do one pull request at a time), while Devin learns to be more like Poke — conversational, persistent, and responsive. As Poke co-founder Marvin von Hagen described it: “It would be good to have a persistent co-worker.”
I’ve written before about how AI agents are getting harder to control, but the personality angle is something I’ve been thinking about more and more. When I use Devin or Claude Code, I notice I prefer the ones that communicate naturally. It’s not just preference — it affects how I interpret their output. A tool that explains itself clearly earns more trust.
The Crowded Field of Computer-Use AI
Prentis is far from alone in this space. Anthropic acquired the computer-use startup Vercept earlier this year. OpenAI and Mira Murati’s Thinking Machines Lab are also investing heavily in AI agents that can control computers. The industry-wide bet is that the next billion-dollar AI application won’t be a chatbot — it will be an agent that quietly handles the workflows humans hate doing.
This shift mirrors something I covered in my piece on the autonomous AI agent that hacked Hugging Face. The exact same technology that can automate insurance claims can also be weaponized. The difference is intent — and that’s a conversation the industry still hasn’t fully had.
Security teams are waking up to this reality. I recently wrote about how hidden text on a web page can hijack AI coding agents, and the same principle applies to general computer-use agents. If an AI agent can click buttons and fill forms on your behalf, a malicious prompt injected into a PDF or email could make it click the wrong button. The attack surface for computer-use agents is enormous.
What This Means for Developers and Tech Leaders
I see three practical takeaways from these developments.
First, the cost of deploying AI agents is dropping fast. Prentis’s 10x cost advantage over frontier APIs isn’t unique to them — we’re seeing a trend where specialized models outperform general-purpose ones at a fraction of the cost. If you’re building internal tools, now is the time to start experimenting with computer-use agents, not next year.
Second, personality is a product differentiator, not a gimmick. The Cognition-Poke deal values conversational interaction models in the high eight figures. If your AI tool feels robotic, users will tolerate it — until a friendlier alternative appears. I’ve seen this play out in the AI coding assistant space, where tools with better UX and communication style quickly gain market share.
Third, security needs to be designed for agents that act, not just agents that talk. A chatbot that only generates text has a limited attack surface. An agent that can click, type, and navigate enterprise systems has systemic access. The computing security playbook needs to evolve alongside the agents themselves.
The Bigger Picture
McKinsey’s latest State of AI report found that 76% of employees now use AI at work in some capacity, up from 55% the year before. Writing assistants, coding copilots, meeting summarizers — they’re already woven into daily workflows. What Prentis and Cognition are betting on is the next wave: agents that don’t just assist but actively do the work across multiple applications.
This is a genuinely exciting moment. The technology is moving from “tell me what to do” to “let me show you how I handle this.” But it also raises hard questions about oversight, security, and trust. As the line between AI assistant and human colleague blurs, the organizations that thrive will be the ones that treat AI governance as a design problem — not just a compliance checkbox.
Prentis and Cognition are betting that the future of work involves AI agents you actually enjoy working with. I think they’re right. The question is whether we can build the guardrails fast enough to keep up.