Software at the speed of agents
Airbnb says AI writes 60 percent of its new code, and agents built a compiler in two weeks. What is possible now, and what still needs people.
A year ago the argument was whether AI could write useful code at all. In 2026 the questions are how much of it to hand over, how fast the result can ship, and who checks it. Here is what companies have said and shown in the last months, and where people still make the difference.
What companies say
On its first quarter call in May, Airbnb said that 60 percent of the code its engineers produced in the quarter was written by AI. In August its chief executive, Brian Chesky, said the time from concept to launch had fallen by as much as 60 percent on some key initiatives, and that Airbnb had shipped nearly 80 percent more features and improvements than in the same six months a year before.
In February Spotify’s co chief executive, Gustav Söderström, said the company’s best developers had not written a single line of code since December. They work through an internal system called Honk, built on Claude Code: an engineer on the way to work asks Claude to fix a bug, receives a new build of the app in Slack and merges it before reaching the office. Cognition, the maker of the coding agent Devin, says Devin wrote 89 percent of the code its own engineers committed, and agents in its Windsurf editor wrote the rest.
For scale, in April 2025 Microsoft’s Satya Nadella put the share written by software in Microsoft’s repositories at 20 to 30 percent, and Sundar Pichai said AI generated more than 30 percent of Google’s code. The figures are the companies’ own, and none of the reports explains how they were measured.
What agents have built
Two experiments show the upper end. In February Nicholas Carlini at Anthropic let 16 instances of Claude work in parallel on a C compiler, written from scratch in Rust. Over nearly 2'000 sessions and about two weeks, for just under 20'000 dollars in compute, they produced 100'000 lines of code that can build the Linux kernel on three processor architectures. It is not a drop in replacement for a production compiler, and its output is slower. Carlini’s warning is the useful part: it is easy to see tests pass and assume the job is done, when that is rarely the case.
In January Cursor described running hundreds of agents on one project. In close to a week they wrote a web browser from scratch, more than a million lines across about 1'000 files. A large migration inside Cursor’s own code took more than three weeks and, in the team’s words, still needed careful review.
The market for building
The money follows. Cursor passed 2 billion dollars in annualised revenue this spring, and in June SpaceX agreed to buy it for 60 billion dollars in stock. Anthropic said in March that run rate revenue from Claude Code had passed 2.5 billion dollars. Lovable, where people describe an app and the platform builds and hosts it, crossed 600 million dollars in annualised revenue this week, and the apps built on it get close to a billion visits a month.
The effect is visible in the app stores. Appfigures counted 60 percent more new apps in the first quarter of 2026 than a year earlier across Apple’s App Store and Google Play, and 80 percent more on iOS alone. TechCrunch’s working explanation is AI coding tools that let people with an idea build their first app.
What still needs people
The bottleneck has moved from writing to checking. In March Anthropic launched a review tool for Claude Code because, as Anthropic’s Cat Wu put it, Claude Code had dramatically increased code output and review could not keep up. Open source maintainers describe the same flood from the other side. The head of VideoLAN, which makes the VLC player, called the merge requests from junior contributors abysmal, and the cURL project stopped its bug bounty after being swamped by what its creator called AI slop.
Security does not come with the speed. This week UpGuard reported about 16'000 databases hosted on Supabase with personal data open to the internet, and TechCrunch linked the problem to apps built quickly with AI. Supabase says its platform is secure by default and that configuration is the customer’s responsibility. Costs need watching too: Uber used up its annual AI budget in about four months and now caps each employee at 1'500 dollars a month per coding tool.
And speed is not automatic. In a study by METR in 2025, experienced open source developers expected AI to make them 24 percent faster and were 19 percent slower, though the authors warned against reading too much into one study. Cognition’s Scott Wu, whose company sells the agents, still says it should always be up to the human what to do.
That is where people stay: deciding what is worth building, how the parts fit together, what may touch customer data, and when something is good enough to ship. The agents write the lines. Architecture, security, taste and review need someone who will answer for the result.
How we work, and what a company can do
We build this way ourselves. Werk, our product for apps and digitalisation, builds web, iOS and Android apps with AI on our own framework and platform. In Studio, every publish is checked, built once, tested, made live, checked again and rolled back on its own if anything fails, so speed never reaches a customer untested. Twin brings a company’s knowledge and rules into Codex for the people who build.
For a company, five steps follow. Write down what you want before an agent writes code. Put tests and review gates in front of every release. Have someone who knows security look at anything that touches personal data. Give agents a budget and measure what ships, not how many lines were written. And keep architecture and the decision to release with named people.
