Cognition’s acquisition of Poke is the kind of deal that looks, at first glance, like a straightforward talent-and-technology purchase. But the more interesting story is what Cognition is choosing to buy: not just an AI capability, but an interaction philosophy—an “AI personality” and conversational model that changes how people feel while they work with an agent.
Poke, the AI assistant many users describe as something you text like a friend, has been known less for raw technical novelty than for its tone, pacing, and the way it handles back-and-forth. It doesn’t simply answer; it collaborates in a way that feels continuous and human-adjacent. Now Cognition is bringing that approach into Devin, its coding agent. The reported valuation places the deal in the low nine figures, which signals confidence without implying a massive consolidation. In other words: this isn’t a “buy the whole category” move. It’s a targeted bet on a specific competitive lever.
That lever is becoming clearer across the AI industry: model quality matters, but assistant behavior is increasingly what determines whether people stick around long enough to benefit from that quality.
For years, the AI conversation has centered on benchmarks—accuracy, reasoning, coding correctness, latency, cost per token. Those metrics still matter. Yet anyone who has used multiple assistants knows that two systems can produce similar outputs while one feels dramatically easier to work with. The difference often comes down to interaction design: how the assistant asks questions, how it confirms assumptions, how it handles uncertainty, how it responds when you change your mind, and how it maintains context without becoming confusing or overly formal.
Poke’s value proposition has been precisely that: a conversational style that reduces friction. Cognition’s decision to integrate that style into Devin suggests the company believes friction is now the bottleneck for adoption of AI agents—not just intelligence.
Devin is already positioned as a coding agent rather than a chat bot. That distinction matters. A coding agent isn’t merely responding to prompts; it’s executing tasks, iterating on code, and coordinating steps that span time. In practice, that means the user experience is not a single conversation—it’s a workflow. The assistant must manage transitions: from planning to implementation, from debugging to verification, from “here’s what I think” to “here’s what I changed.” If the assistant’s communication style doesn’t match the user’s mental model of the work, the user spends their attention correcting the process instead of guiding the outcome.
This is where Poke’s interaction model becomes more than branding. A friendly, text-like conversational approach can do something subtle but powerful for agentic systems: it can make iterative collaboration feel natural. When an agent needs to ask clarifying questions, a conversational assistant can phrase those questions in a way that invites cooperation rather than demands precision. When an agent makes a mistake, the way it acknowledges the error—and the way it proposes the next step—can determine whether the user trusts the system enough to continue.
Cognition’s move implies that Devin’s future advantage may come from improving the “loop” between human intent and agent execution. The loop is where most real-world failures happen. Even if the agent can generate correct code in isolation, it can still fail the workflow if it doesn’t communicate effectively during the iteration cycle.
Consider what happens in typical coding work. Developers rarely specify everything up front. They explore. They try an approach, see what breaks, adjust requirements, and refine. A good assistant mirrors that exploration. It doesn’t just deliver a final answer; it supports a sequence of decisions. That requires a conversational layer that can handle ambiguity gracefully.
Poke’s reputation for being easy to talk to suggests it has learned how to keep that layer from becoming either too robotic or too vague. Friendly doesn’t mean careless. It means the assistant can maintain momentum while still being accountable. For an agent like Devin, that balance is crucial. If the assistant is overly formal, users may feel like they’re submitting tickets rather than collaborating. If it’s overly casual, users may feel uncertain about what the agent is doing and why.
The acquisition also reflects a broader shift in how companies think about “personality” in AI. For a long time, personality was treated as a surface feature—something you could add after the core model worked. But the industry is increasingly recognizing that personality is not only about tone; it’s about policy. It’s about how the system chooses what to say, when to say it, and how to structure the interaction so that the user can steer the agent effectively.
In other words, personality is becoming a product interface for control.
When you text a friend, you don’t need a manual for every nuance. You can correct them midstream. You can ask for clarification without feeling like you’re breaking a protocol. You can be slightly messy, and the other person still understands the intent. Translating that dynamic into AI agents is difficult because it requires more than language generation. It requires consistent interaction patterns: how the assistant confirms goals, how it summarizes progress, how it handles partial information, and how it recovers when the conversation drifts.
Cognition’s decision to bring Poke’s interaction model into Devin suggests it wants Devin to behave less like a tool that waits for perfect instructions and more like a collaborator that can work with imperfect ones. That’s a meaningful shift for coding agents, which often struggle with the “last mile” of usability: the moment when the user realizes they need to guide the agent’s understanding of the task.
There’s also a strategic reason this acquisition makes sense now. As AI coding capabilities become more widely available—through open models, cloud APIs, and competing agent frameworks—the differentiation based purely on raw coding performance becomes harder to sustain. If multiple systems can generate plausible code, the winner is likely to be the one that helps users complete tasks with fewer interruptions, fewer retries, and less cognitive overhead.
Communication style directly affects those outcomes. A system that communicates clearly reduces the number of times a user has to ask, “Wait—what did you do?” or “Why did you choose that approach?” A system that asks the right questions early reduces the number of times the user has to backtrack later. A system that maintains context and explains changes reduces the risk of silent failure.
In that sense, Poke’s “friendly” interaction model may be less about charm and more about operational clarity. Friendly can be a delivery mechanism for structured thinking.
One unique angle here is how this acquisition reframes what “competitive advantage” means in AI. Many companies still treat the model as the engine and the UI as a wrapper. Cognition appears to be treating the conversational layer as part of the engine. That’s a subtle but important shift. It suggests that the company sees interaction design as a form of intelligence—because it shapes the feedback the system receives and the decisions it makes.
Agentic systems learn from interaction in a practical sense even if they aren’t training online. They adapt through prompting, through the user’s corrections, through the conversation’s evolving constraints. If the assistant’s communication style encourages better feedback, it can improve outcomes over time. Even without parameter updates, the system can become more effective because the user’s guidance becomes more precise and more frequent.
This is why the acquisition is notable beyond Cognition. It’s a signal to the market that conversational design is not a “nice-to-have.” It’s a core component of agent performance in the real world.
There’s another implication: the industry may be moving toward modular agent stacks where interaction models are portable. Today, many agent systems are tightly coupled: the model, the tool use, the memory strategy, and the conversation policy are all intertwined. But acquisitions like this hint at a future where companies can swap in better conversational policies the way they swap in better retrieval components or improved planning modules.
If Poke’s interaction model can be integrated into Devin, it suggests Cognition believes the conversational layer is sufficiently modular to be transplanted. That would accelerate iteration. Instead of reinventing the entire interaction approach from scratch, Cognition can incorporate proven patterns and then tune them for coding workflows.
That tuning is likely where the real work will happen. A texting assistant and a coding agent operate in different contexts. Texting is often about emotional resonance, quick back-and-forth, and maintaining a sense of continuity. Coding agents must also manage technical correctness, tool execution, and multi-step tasks. The conversational style must therefore be adapted so that friendliness doesn’t obscure technical details.
For example, when Devin proposes a plan, it needs to communicate tradeoffs: what it will do first, what assumptions it’s making, and what it expects to find. When it edits code, it needs to show diffs or explain changes in a way that helps the user verify correctness. When it encounters errors, it needs to interpret logs and propose debugging steps. A “friend-like” tone can help users stay engaged during these moments, but the assistant still has to be rigorous.
The best version of this integration would likely look like a hybrid: conversational warmth paired with engineering discipline. The assistant should feel approachable while still being explicit about uncertainty and next actions. It should ask clarifying questions without sounding interrogative. It should summarize progress without being vague. It should recover from mistakes without blaming the user or pretending nothing went wrong.
If Cognition pulls that off, Devin could become more than a coding agent that generates code. It could become a coding partner that reduces the friction of iteration—making it easier for developers to delegate parts of the workflow without losing control.
This is also where the “low nine figures” valuation becomes meaningful. It suggests Poke is not being acquired for a massive proprietary dataset or a foundational model. It’s being acquired for something more specific: the interaction model, the product experience, and the know-how behind building an assistant that people want to talk to. That kind of asset is hard to replicate quickly. You can train a model, but you can’t easily manufacture trust and conversational fluency without extensive product iteration.
Trust is the hidden currency in agent adoption. Users don’t just want correct