For the past two years, the AI coding wars ran on benchmarks. Who topped SWE-bench. Who shipped the fastest agent. And who closed a pull request with the fewest human interventions. This week, Cognition — the company behind autonomous coding agent Devin — signaled a different fight ahead. The next edge won’t come from a leaderboard. It’ll come from how the agent actually talks to you.
On July 23, Cognition acquired The Interaction Company of California, maker of Poke, an AI assistant people text like a friend over iMessage, SMS, Telegram, and, in some markets, WhatsApp. The price sits in the low nine figures. It’s Cognition’s third acquisition in roughly a year, following the 2025 purchase of agentic IDE Windsurf and the July 20 acquisition of post-deployment automation startup TierZero. Together, the three deals sketch a company chasing the entire lifecycle of software work: write it, ship it, keep it running, and now, talk about it like a colleague would.
The Product Nobody Was Benchmarking
Poke never tried to win a coding benchmark. It isn’t a coding tool at all. Its real proof of value came down to something simpler: people kept talking to it. In the three months before the acquisition, Poke users exchanged more than 100 million messages with it, according to Cognition. That’s a striking figure for a product that only launched in March 2026 and still counts hundreds of thousands of users.
A benchmark score doesn’t explain that number. An engagement signal does — and it’s a number most coding-agent companies rarely get to cite. Most still sell a tool: write the code, fix the bug, open the pull request, move on. Poke sold something closer to a standing relationship. It checked in proactively, cracked the occasional joke, and behaved as if it remembered what you asked yesterday.
Interaction co-founder Marvin von Hagen put the thesis plainly in comments to TechCrunch. People generally prefer colleagues with personality over ones that behave like robots, he said — and a co-worker who happens to be a software engineer can still land a joke you’ll actually find funny.
Poke also isn’t alone in betting that a messaging app, not a dashboard, is the right interface for an AI agent. Open-source projects like OpenClaw have built similar agents around WhatsApp, Telegram, and Discord, suggesting the “text it like a friend” model is becoming a category, not just one startup’s gimmick.
Buying the Interaction, Not Just the Model
What makes this deal unusual is the direction of value flow. Cognition didn’t acquire Poke for its underlying model. Poke will keep running largely as-is through the end of 2026. Next year, Cognition plans to experiment with routing some Poke tasks through its own SWE-1.7 model, though it hasn’t committed to a full integration. What Cognition actually bought is the interaction layer — the accumulated product decisions about tone, memory, proactivity, and pacing that make an agent feel less like autocomplete and more like a co-worker.
That layer is harder to copy than a system prompt. It represents months of real-world feedback: how proactive an assistant should be, how much personality is too much, when a joke helps instead of grating.
A prior relationship also sits behind the deal, which makes it less of a cold acquisition than it looks on paper. Cognition co-founder Scott Wu said he and co-founder Walden Yan had been angel investors in Interaction. Wu also noted on X that von Hagen was on set for Devin’s original 2024 launch video, framing the acquisition, in his words, as “two years in the making.”
The Real Prize: A Devin That Remembers
The near-term roadmap stays incremental. Poke remains independent through the rest of the year and stays live on Apple’s Messages for Business platform — a milestone in its own right, since Poke became the first third-party AI agent approved to run there in June. The longer-term ambition, as both von Hagen and Wu described it, points somewhere more interesting.
Devin currently handles one pull request at a time. Von Hagen suggested Poke could eventually orchestrate multiple Devin sessions at once, coordinating tasks the way a technical lead runs several workstreams in parallel while helping Devin retain context across sessions. That’s a step beyond a single request-response agent and closer to the kind of multi-agent AI orchestration that routes, sequences, and monitors several tasks at once rather than running one linear chain.
“It would be good to have a persistent co-worker,” von Hagen said.
There’s a Catch: Personality Isn’t Free
Poke’s own numbers hint at the tension underneath this bet. High engagement, hundreds of thousands of users — and, by von Hagen’s own account, expensive to run and hard to turn a profit on. Agents built around frequent conversation, memory, and unprompted check-ins likely carry a different cost structure than a transactional coding tool. Every extra exchange adds to the inference bill, and those costs add up fast at scale for any company running agents this chatty.
Cognition is betting that folding Poke into its own infrastructure fixes the unit economics without sanding off the personality that made people stick around. Get that wrong, and this deal becomes the cautionary tale about why coding tools stayed robots.
The Bet Cognition Is Actually Making
Either way, judging AI agents purely on what they can do just picked up a second axis: what they’re like to work with. Cognition is betting that axis can matter nearly as much as raw capability. Not because Devin needs better jokes, but because the next bottleneck in agentic software may not be intelligence at all. It may come down to whether people trust the thing enough to keep delegating to it.
Related: GLM-5.2 Just Matched Claude Mythos — And Anyone Can Download It
