Why Cognition bought Poke: AI personality is becoming a competitive advantage

Cognition did not buy Poke for its user base. It bought the thing that made those users stay: a personality that feels human enough to joke with. The price — low nine figures — is practically a rounding error for a company building a coding agent that raised at a $2 billion valuation. But the signal is loud. The Interaction Company cracked something most AI labs treat as an afterthought. They made an assistant that replies with slang, humor, and the irregular cadence of a friend. Not a persona. A personality.

The distinction matters. Persona is skin. Personality is structure. Poke's agent remembers context across weeks, initiates follow-ups unprompted, and matches the user's tonal register. It does not wait for instructions. It collaborates. That is what Cognition wants inside Devin. Scott Wu put it plainly: working with Devin should feel like working with a colleague who happens to be software. Colleagues make jokes. They push back. They know when you're tired.

Marvin von Hagen, Poke's co-founder, framed it as coworker dynamics. You prefer colleagues with personality over robots. The market agreed. Hundreds of thousands of users exchanged over 100 million messages in three months. They used Poke for travel, health, finance, scheduling, education. But the dominant use case was productivity — email triage, reminders, to-dos. The boring stuff. The stuff you dump on a human assistant because you trust them to handle it without supervision. That trust came from interaction quality, not model benchmarks.

Here is the uncomfortable truth for pure-play model companies: the best model loses if the interface feels like a terminal. OpenAI knows this. Anthropic knows this. Google knows this. But they solve it with system prompts and RLHF patches. Poke solved it by building the interaction model first, then wrapping the LLM around it. The personality is not a veneer. It is the architecture.

That architecture was expensive. Von Hagen admitted Poke struggled to turn a profit. Inference costs for a chatty, proactive, memory-heavy agent running across iMessage, SMS, Telegram, and WhatsApp crushed unit economics. Apple's Messages for Business approval in June was a distribution win, not a margin fix. Cognition's infrastructure — custom model serving, optimized routing, the whole stack they built for Devin — changes the math. Poke gets cheaper. Devin gets warmer.

The angel investment history is not coincidence. Wu and Yan funded von Hagen's team before they competed. They watched the interaction layer compound in value. They saw retention curves that model improvements alone could not explain. Then they bought the source.

Next year the experiment starts in earnest. Poke will test Cognition's newest models. Devin will test Poke's interaction primitives. The boundary between consumer assistant and coding agent will blur. A developer who texts Devin like Poke — slang, shorthand, half-formed thoughts — and gets working code back is a developer who stops context-switching. That is the product.

Skeptics will call this a acqui-hire with a brand attached. They are wrong. The Interaction Company's IP is not the team. It is the dataset of human-AI rapport: millions of conversations where users treated the agent as a peer. That dataset trains interaction models no benchmark captures. Cognition now owns it.

The broader lesson: model commoditization is real. Interaction differentiation is the moat. Companies that treat personality as a feature will rent it. Companies that treat it as infrastructure will own it. Cognition just chose ownership.