The AI arms race just shifted from better models to better manners.
The Summary
- Cognition acquired Poke, the text-it-like-a-friend AI assistant, for low nine figures to bring conversational personality to Devin, its AI coding agent
- Interface design and interaction patterns are emerging as competitive moats in a world where everyone has access to similar foundation models
- The bet: developers will choose agents they actually want to work with, not just ones that technically work
The Signal
Cognition didn't need Poke's model. They needed its vibe. The acquisition signals something the AI infrastructure layer has been slow to acknowledge: when the underlying models commoditize, the wrapper becomes the product.
Devin made headlines as an autonomous coding agent that could tackle full software engineering tasks. But autonomous doesn't mean pleasant to work with. Early Devin users reported a common pattern: the agent worked, but felt like delegating to a brilliant intern who never asked clarifying questions and never picked up on context. Poke solved the opposite problem. It couldn't code, but people texted it constantly because it felt like talking to someone who got you.
"When foundation models converge in capability, personality becomes the moat."
The convergence thesis is playing out faster than anyone predicted. GPT-4, Claude 3.5, Gemini Ultra — the performance gaps are narrowing. What separates ChatGPT from Claude isn't raw intelligence anymore. It's whether the response feels like it came from a careful librarian or a eager teaching assistant. That's interface design, not model architecture.
Cognition is betting that developers want coding agents that feel less like command-line tools and more like pair programmers. The Poke acquisition brings:
- Conversational memory that persists across sessions without feeling creepy
- Interaction patterns that mirror text-based collaboration, not robotic task completion
- UI/UX design that treats the agent as a colleague, not a compiler
This matters because agent adoption has a personality problem. Enterprise buyers evaluate agents on benchmark performance. Individual users choose agents they don't hate working with. Cognition is building for the second group, because they're the ones who'll actually use the tool daily.
The low nine-figure price tag — likely $100-150 million based on industry norms for "low nine figures" — suggests Cognition sees personality as worth more than incremental model improvements. For context, that's acquisition pricing usually reserved for companies with significant revenue or unique technical IP. Poke had neither. It had taste and interaction design that made people choose it over free alternatives.
The Implication
Watch for more personality acquisitions as the agent economy matures. The companies that win won't just build capable agents. They'll build agents people prefer. That's a design problem, not a training problem.
For teams building agents: benchmark performance gets you in the room. Interaction design keeps users coming back. If your agent passes every eval but nobody wants to work with it, you've built a tool, not a teammate.