Cognition buys Poke for ~$100M to give its AI coding agent Devin a personality

Cognition, the startup behind the AI coding agent Devin, paid roughly $100 million for Poke, a consumer AI assistant people text like a friend across iMessage, SMS, Telegram, and WhatsApp. The deal, disclosed in a TechCrunch report on Thursday, and interaction design into Cognition's roadmap. Poke cleared 100 million messages in three months and still could not make money. The bet is that Devin can subsidize the interaction layer that broke Poke as a standalone business.

The product experience is closer to texting a sharp friend than commanding a tool. Poke launched in March 2026, and its co-founder Marvin von Hagen told TechCrunch the service has handled more than 100 million messages in the past three months from hundreds of thousands of users. Use cases span travel, health, finance, scheduling, education, and basic productivity, with email and reminders bundled in. It runs on every major messaging surface, including WhatsApp in some markets.

Cognition's argument for the price is not that Poke works as a standalone business. Von Hagen conceded the service was expensive to run and difficult to profit. The argument is that Poke has something Devin needs: a personality layer, a slang register, a habit of initiating, a way of phrasing suggestions so a user accepts them. Cognition plans to graft Poke's interaction model onto Devin, the autonomous coding agent that competes with Cursor, Claude Code, and a fast-growing field of agentic developer tools. Coding agents are getting similar at the model layer. Cognition is buying a way to stand apart at the surface.

The economics underneath that bet are the part of the story that should make readers pause. One hundred million messages from hundreds of thousands of users in three months is real adoption, and the figure comes from Cognition itself rather than independent measurement. It also produced no profit. If personality is the moat, the first product to commercialize it at scale should already be paying its own way. Poke was not. The interaction design Cognition bought is unproven as a business. The only proof point on offer is user pull, and user pull did not solve the unit problem.

That gap is the test the deal sets up. Cognition does not need Poke to be a profitable consumer chatbot. It needs Poke to be a personality engine it can ship inside a coding agent that already has paying customers. Devin's existing pricing, the willingness of engineering teams to pay for an autonomous coder, and the gross margin on agent compute all become the subsidy that consumer AI could not find. If Devin's revenue model can absorb the interaction-layer cost that broke Poke, the moat thesis holds. If not, Cognition will have paid nine figures for tone of voice.

A relationship fact belongs on the table. Cognition co-founders Scott Wu and Walden Yan were angel investors in the Interaction Company before the acquisition. That is a disclosed prior financial tie, and it shapes how to read the price. The people buying had already put personal money in, and the deal consolidates a position rather than arms-length discovery.

The falsifier is worth naming too. If a frontier model lab ships personality features directly in the model API, or if Poke-style interaction patterns get cloned the way chatbot UIs have been, the interaction layer stops being a moat and becomes table stakes. The first major coding agent to ship a comparable tone without paying for it will be the evidence the bet was wrong. Watch the model release notes. That is where this thesis is most likely to break.

For now, Cognition is the first major AI coding-agent company to publicly stake nine figures on the idea that interaction design, not raw model capability, decides who wins in agent products. Poke's three months of 100 million messages and zero profit are both halves of that bet. The product is the proof. The missing profit is the question.