The wildest 72 hours
On Friday, July 11, 2025, Google announced a $2.4 billion deal to poach Windsurf's CEO Varun Mohan, co-founder Douglas Chen, and several research leaders. The move was a reverse acqui-hire: Google got the talent and a nonexclusive technology license, and Windsurf's 250-person team was left without its top leadership. Hours earlier, OpenAI's $3 billion acquisition offer for Windsurf had expired. The company was adrift.
Scott Wu, Cognition's CEO, made the first call after 5 PM that Friday. By Monday morning, he had a signed agreement. Jeff Wang, Windsurf's interim CEO who had been thrust into the role days earlier when Google hired Mohan, called it "the wildest rollercoaster ride of my career." In a post on LinkedIn, he wrote: "To our new teammates at Cognition: we at Windsurf feel incredibly lucky to be joining a team that shares our vision, our deep commitment to our users, and, most importantly, our values."
Cognition acquired Windsurf's IP, product, trademark, brand, and all remaining employees. Every Windsurf employee participated financially in the deal, with vesting cliffs waived for their work to date. The deal brought $82 million in ARR, 350+ enterprise customers, hundreds of thousands of daily active users, and world-class go-to-market and engineering teams. Three days later, they shipped Wave 11. They called it "Just Keep Shipping."
"The last 72 hours have been the wildest rollercoaster ride of my career. To our new teammates at Cognition: we feel incredibly lucky to be joining a team that shares our vision."
What they built in twelve months
The year since the acquisition has been one of the most aggressive shipping cadences in the history of software. Cognition went from a single-agent product to a multi-surface, multi-model platform. They built their own models: SWE-1.5 at ~1000 tokens per second, then SWE-1.6, then SWE-1.7. They launched Devin Review, Devin CLI, Devin 2.2, and Cognition for Government. Windsurf 2.0 unified every agent into a single command center. Devin Desktop unified the brand. They put their money where their mouth is with the AI Productivity Guarantee. They became a global company with offices in San Francisco, New York, London, Singapore, and Tokyo.
The numbers that matter
Let's be precise about what "10x" means here. In May 2025, Cognition's annualized revenue run rate was $37 million. By May 2026, it was $492 million. That's a 13x increase in twelve months. Enterprise usage of Devin grew more than tenfold in the first months of 2026, with roughly 50% month-over-month growth sustained for six consecutive months. The company raised over $1 billion in Series D funding at a $26 billion valuation, more than doubling its previous $10.2 billion mark in just eight months. The team grew from 44 to 350 people. Twenty million lines of code were written locally by the team. And 89% of the code committed by Cognition's own engineers is committed by Devin.
The customer list reads like a Fortune 500 roll call: Goldman Sachs, Citi, Mercedes-Benz, Dell, Santander, Palantir, NASA, and units of the U.S. Army and Navy. Mercedes-Benz compressed an eight-month legacy modernization project to eight days. Itau, Latin America's largest bank, resolves 70% of its security vulnerabilities automatically through Devin. Systems integrators Infosys and Cognizant have embedded Devin into client delivery workflows. Japan's Sapporo city government modernized over one million lines of legacy code in roughly a quarter of the estimated time using Devin.
The swarm: beyond single agents
On July 1, 2026, Cognition launched Devin Security Swarm, and it represents something fundamentally new. The press release called it a system "to help security teams find exploitable vulnerabilities, validate them at runtime, and fix them." But the architecture underneath is the real story.
They call it Agentic MapReduce. The name borrows from the decades-old distributed computing pattern that made processing datasets too large for one machine tractable. Cognition applied that logic to reasoning, not just data processing. A single agent cannot hold an entire enterprise codebase in its working memory, so it either scans shallow or scans narrow. Agentic MapReduce splits the difference.
The architecture has five phases. First, a planner agent studies the repository and builds a threat model, generating deterministic selectors that surface candidate locations. Second, a swarm of investigation agents runs in parallel, each reasoning over a bounded shard of the codebase. Third, a reducer agent combines all findings, deduplicates, and composes cross-file attack chains. Fourth, each serious finding is reproduced in an isolated sandbox against a running build to confirm exploitability. Fifth, Devin writes the patch and opens a remediation PR.
On a benchmark of 50 real-world vulnerabilities tied to published GitHub Security Advisories across 14 languages, Security Swarm found 36. That's 72% recall, the highest of any tool tested, at roughly two-thirds the cost of the next most accurate alternative. Three critical vulnerabilities were found exclusively by Devin and missed by every other tool tested.
Single agents hit a context wall. A single agent cannot hold an entire enterprise codebase in its working memory. It either scans shallow or scans narrow. Agentic MapReduce solves this by fanning out focused agents on bounded shards, then reducing the results into a single coherent report.
Verification separates this from every other scanner. Most AI code-scanning tools flag pattern matches and leave triage to a human. Security Swarm reproduces each exploit in an isolated sandbox before marking it confirmed. A tool that reports fifty possible vulnerabilities and confirms three earns more trust than one that reports three hundred and confirms none.
The pattern is reusable infrastructure. Cognition published extensive documentation on the Agentic MapReduce architecture, signaling they see it as reusable beyond security. The same map-reduce pattern could scale to any whole-codebase task where a single context window is the bottleneck.
The vision: from agents to swarms
What Cognition understands, and what the rest of the industry is still catching up to, is that the future of AI is not a single agent doing one thing well. It's swarms of specialized agents coordinating on problems too large for any individual context window. The Security Swarm is the first production deployment of this thesis, but the implications extend far beyond security.
Consider what this means for the trajectory of AI. The first generation of AI coding tools was autocomplete. The second generation was single agents that could handle isolated tasks. The third generation, which Cognition is now defining, is swarms of agents that can reason across entire codebases, validate their findings against running systems, and remediate what they discover. The jump from "agent that writes a function" to "swarm that secures an enterprise" is not incremental. It's categorical.
Scott Wu has said that the future of AI coding will not depend on a single large language model but on systems capable of intelligently combining multiple models and tools together. The Security Swarm is the proof of concept. If a swarm of agents can reason across an entire enterprise codebase to find security vulnerabilities, validate them at runtime, and write the patches, then the same architecture can reason across that codebase for performance optimization, architectural review, dependency analysis, or feature planning. The context window is no longer the bottleneck. The swarm scales with the problem.
The culture that made it possible
None of this happened by accident. Cognition runs what CEO Scott Wu has publicly described as an "extreme performance culture." The team typically arrives before the 11:30 AM daily all-hands and often stays past midnight. Weekend work is routine, not exceptional. Founding engineer Adhyyan Sekhsaria noted the team had worked until 3 AM the day before his interview. After acquiring Windsurf, Wu offered nine months of severance to anyone unwilling to commit to the pace. It goes, as reporting has noted, "even further than 996."
This is not a culture everyone wants to join. But it is a culture that ships. And in twelve months, it shipped more than most companies ship in a decade. The team wrote 20 million lines of code locally. They launched proprietary models, multiple products, expanded to five countries, grew revenue 13x, and pioneered a new architectural paradigm for AI agents. The "Just Keep Shipping" ethos from Wave 11 was not a slogan. It was an operating model.
"Among all the teams in the AI space, Cognition was literally the one we have respected the most, and they are a perfect fit to bring Windsurf to the next phase."
The shoutout
To the Cognition team: you lost your would-be partners' CEO to Google in a $2.4 billion heist. OpenAI's $3 billion acquisition offer had just expired. The 250-person Windsurf team was leaderless and adrift. You had 72 hours to make a deal. You made it on Monday morning.
Then you did the hard part. You integrated two companies with different cultures, different products, and different identities. You unified them under one brand. You built proprietary models. You launched a desktop app, a CLI, a review tool, and a government product. You expanded to five countries. You grew revenue from $37 million to $500 million. You grew the team from 44 to 350. You wrote 20 million lines of code. And you shipped Devin Security Swarm, the first production deployment of Agentic MapReduce, proving that the future of AI is not single agents but coordinated swarms.
You bet on yourselves. You were right.
As Cognition said in their anniversary post: "Engineers are becoming super engineers, and people who never coded before are able to leverage themselves in ways they never imagined." The swarm era is here. And the team that built it did so while absorbing a company that had just lost its CEO, integrating it in a weekend, and shipping three days later.
Just keep shipping.
Cognition AI — applied AI lab, makers of Devin, the first AI software engineer. Founded 2023 by Scott Wu, Steven Hao, and Walden Yan. cognition.com