Cognition.
The lab that ships.

01 / NOWAn engineer that never sleeps

On a Tuesday morning in 2026, somewhere in a sandboxed VM, a software engineer named Devin is fixing a bug it found on its own. It has read the issue, traced the stack, written a patch, run the tests, opened a pull request and pinged a human for review. The human is in a different time zone. Devin does not mind.

The engineer works for Cognition, a four-floor office in San Francisco where roughly 200 humans build the agents that, increasingly, build everything else. They started in 2023 with a thesis that sounded reckless and now sounds, if anything, conservative - that software engineering is a reasoning problem, and reasoning is something machines are about to get very good at.

The proof, for now, is in the receipts. $400 million raised in September 2025 at a $10.2 billion valuation. Annualized revenue from Devin went from $1 million to $73 million in nine months. In April 2026, reports surfaced that the company was in talks for a round at $25 billion. Numbers that look like typos until you check them twice.

02 / THE PROBLEMSoftware was always the bottleneck

The premise sounds obvious once you say it out loud, which is the surest sign that nobody had built a company around it yet. Every company is becoming a software company. Most of them are bad at software. The ones that are good at software still cannot hire enough engineers to do half the things they would like to do. Backlogs are infinite. Pull request queues are theology. Roadmaps are aspirational fiction.

The conventional response was to make tools that help humans type faster - linters, autocomplete, copilots. Useful, in the way a better stenographer is useful to a novelist. It did not fix the bottleneck. The bottleneck was thinking.

So Cognition picked the harder bet. Not a copilot. An agent. Something that could be given a high-level goal - "fix this flaky test", "add SSO to the dashboard", "investigate why the queue lags on Mondays" - and would go away, plan, code, test, browse documentation, and come back with a verdict. The wager was that the model of the future is not a tool, it is a teammate.

03 / THE BETThree gold medals and a hunch

The three co-founders met where many of the more interesting bets in this industry start - at the International Olympiad in Informatics, the closest thing competitive programming has to the World Cup. Scott Wu won gold three years in a row, placing first in 2014. Steven Hao and Walden Yan both won gold too. Between them, the founding team owns more IOI hardware than the national programs of most G20 countries.

Wu has a twin brother, Neal, who also competed at IOI. Family dinners must be brutal.

The hunch the three of them had in mid-2023 was that the new transformer-based models were close enough to passable reasoning that, given the right scaffolding - a sandbox, tools, memory, the ability to back up and try again - they could perform real engineering work. Not just snippet completion. End-to-end tickets. By March 2024 they had a demo. By the time the world stopped arguing about whether the demo was real, they had revenue.

04 / THE PRODUCTWhat Devin actually does

Strip away the marketing and Devin is, at its core, a virtual machine with three things attached - a shell, an editor, and a browser - plus a large language model that controls them. You give it a task. It plans, it edits files, it runs tests, it reads docs, it asks questions in Slack when it gets stuck. When it is done, it opens a pull request the way a person would.

It is not magic. It is also not subtle. Cognition's engineers say that Devin now writes a meaningful share of Cognition's own code, which is either the most reassuring or the most ouroboric data point in modern software, depending on your mood.

In July 2025, Cognition acquired Windsurf, the agentic IDE whose leadership had just been licensed away to Google in one of the stranger talent deals of the year. The remaining team, the trademark, the customer book - all of it landed at Cognition in roughly thirty-six hours. Today, Windsurf is where developers sit and Devin is who they delegate to. The combined product covers the full surface area of software work, from the first line of an idea to the closing of a ticket.

What you can actually use it for

Customers run Devin against backlogs of small-to-medium tickets, deploy it as a code reviewer that flags problems before humans do, point it at flaky tests until they stop being flaky, and let it pair on migrations - the long, dreary refactors that swallow weeks of senior engineering time. The reported productivity gain depends on which executive you ask, but the customer retention numbers do not lie.

05 / THE PROOFThe chart that does the talking

If you mistrust the rhetoric - sensible posture, in this industry - the numbers are unusually well-behaved. Devin's annualized run-rate revenue went from a rounding error in September 2024 to a real, audited business nine months later. Enterprise adoption doubled quarter on quarter. The customer list reads like a Bloomberg ticker.

Who's buying

Enterprise customers reported across earnings calls and case studies include Goldman Sachs, MongoDB, Nubank, Ramp and a long tail of companies that prefer not to be named, in part because their CFOs are still working out how to account for software written by something that does not draw a salary. Across Devin and Windsurf, hundreds of thousands of developers touch a Cognition product on a given day.

06 / THE MISSIONAn applied lab, not a SaaS shop

Cognition's leadership is careful about a particular phrasing. They are an applied AI lab, not a tool company. The distinction is not vanity. SaaS shops optimize for the present product surface. Labs optimize for the next capability frontier. The former are usually larger; the latter are usually weirder. Cognition is, by most accounts, both.

The internal culture is competitive-programmer-meets-shipped-by-Friday. Small teams. High ownership. A famous willingness to swallow the calendar - the Windsurf deal closed overnight not because anyone wanted to make a movie about it, but because the team did not see a reason to wait.

The mission, in plainer English: build AI teammates capable of taking on real engineering work end-to-end, so the humans on a team can do the things only humans can - decide what should exist, judge whether the result is any good, and apologize to the customer when it is not.

07 / TOMORROWWhy this matters by the end of the decade

If Cognition is correct - and the chart suggests they are at least directionally correct - then the shape of an engineering team is about to change. Less hiring at the junior tier, more delegation, more parallelism, more code, more code review, more code review of code reviewed by code, more managers in the formal sense than the team-size accounting can currently explain.

This is not unequivocally good news. It is also not the end of the profession. The jobs that survive will be the ones that involve taste, judgment, and the willingness to be the person who says no. Those have always been the best parts of software engineering anyway. The drudgery - the test stubs, the upgrades, the migrations, the boilerplate - is exactly the work the founders of Cognition spent their teenage years trying to skip past on a stopwatch.

Back in the sandboxed VM, on that Tuesday morning, Devin is still working. It will be working tomorrow. It will be working on Christmas. Somewhere in San Francisco, three founders who used to compete against the clock are now building the clock itself. The bottleneck has moved. The lab keeps shipping.

08 / WHERE TO GOFind Cognition