AI coding agents have broken the relationship between headcount and software output
Across companies of every size, pull request volumes, productivity multipliers, and build timelines are moving in the same direction at the same time. The evidence now stacks high enough to demand a harder look at what engineering capacity will mean in two years.
Andrew Feldman, who leads Cerebras, describes a shift that took eight months. Token spending per engineer at the company moved from under $1,000 to $25,000 or $30,000. The engineers driving that spend, he says, have gone from being “sort of 10x guys to being 100x guys.” That compression, from a standing start to a hundredfold productivity gain inside a single calendar year, is not an anecdote about one unusually aggressive team. It is a data point that rhymes with what leaders at companies orders of magnitude larger are now reporting.
Patrick Collison, chief executive of Stripe, put numbers to a similar shift. Stripe’s internal agents, which the company calls Minions, were generating 1,200 pull requests per week when Stripe first wrote about them. By last week, that figure had reached 7,000. About 30 percent of all Stripe pull requests now come from agents. Most of that output, Collison adds, traces back to a single engineer who orchestrates 16 agents simultaneously from one screen. A single engineer can now do, in his telling, what two teams of engineers could do two years ago.
At Uber, Brad Gerstner reports that more than 70 percent of pull requests are now attributed to local or cloud agents. Tobi Lütke, chief executive of Shopify, says the figure at his company has crossed 50 percent and is moving higher. Lütke adds a detail that carries its own weight: many of Shopify’s best engineers have not written code since December. “December changed everything,” he says. Wesley Huff relays a secondhand claim that Spotify engineers have not written a line of code since December, a claim Daniel Priestley separately repeats. Both are citing what they heard, not first-party data, and should be read accordingly. The direction of the claim, however, fits every figure reported from inside the companies.
They've gone from being sort of 10x guys to being 100x guys. Andrew Feldman
The gains are not confined to engineering organizations with hundreds of developers and dedicated infrastructure. Andrew Wilkinson describes a person on his team who had never coded in his life, told about Claude Code and initially skeptical. Within about a week, that person had built a working tool. Jason Lemkin describes replacing a $600,000 Salesforce contract with a vibe-coded customer relationship management system built in three weeks. Marc Andreessen puts the broader productivity premium at roughly 20x for leading-edge programmers compared with a year ago. Anton Osika, whose platform Lovable has now seen more than 50 million apps built on it, reports one million new projects created every single week. These numbers span companies, roles, and experience levels. The vector is consistent.
A quieter signal comes from research environments. Ethan He observes that coding models can now implement something in a few hours that previously took weeks, which means compute has become the bottleneck for iterating on ideas rather than the time required to write the code. Eric Jang makes the cost compression explicit: work that once required a whole team of research scientists and millions of dollars of compute can now be done for a few thousand dollars of rented compute. Dylan Patel describes a research project completed by one person that would, by his estimate, have required a team of 200 economists working for a year.
The organizational implications are beginning to surface. Martin Casado, a general partner at Andreessen Horowitz, says agentic coding moved from useful to transformative at the beginning of this year. Evan Spiegel, chief executive of Snap, notes that many of his designers are now shipping code, a role boundary that was firm a year ago. Dara Khosrowshahi, chief executive of Uber, says that within roughly five years, as engineers grow more productive, the company may stop adding engineering headcount and instead add agents and buy more GPUs. Cat Wu describes a threshold crossed only recently, where it became reliable to run multiple code review agents simultaneously across an entire codebase and synthesize issues that actually need human attention before a merge.
The mechanism behind all of this is not subtle. Gavriel Cohen, an open-source maintainer, frames it plainly from the other side: coding agents have made it exponentially easier to open pull requests, creating a triage and review burden that is genuinely hard to keep up with. The friction that once governed how much software a team could produce has moved. What remains is the question of whether the institutions, review processes, and hiring practices built around the old constraint will move with it, or whether the gap between what is now possible and what organizations are organized to absorb will simply keep widening.