AI agents crossed a capability threshold at the start of 2026, and the numbers inside major engineering organizations show it

Martin Casado, a general partner at Andreessen Horowitz, puts the inflection plainly: agentic coding went from being kind of useful to really changing everything at the beginning of this year. That framing matches what the data at several large organizations actually shows.

Patrick Collison, Stripe’s chief executive, offers the most concrete numbers. Stripe’s internal agent system, called Minions, was generating 1,200 pull requests per week when the company first blogged about it in January or February. By the most recent week Collison described, that figure had reached 7,000. About 30 percent of all Stripe pull requests in that period came from Minions. The majority of those, Collison notes, originated from a single engineer who orchestrates 16 agents simultaneously from one screen.

The pattern at smaller scale is equally clear. Walden Yan, who works on Devin, reports that Devin’s commit percentage across its own repositories rose from 16 percent in January to 80 percent in March. Yan is direct about what sits behind that shift: multi-agent systems, he says, “were very much not at all possible a year ago.” One big difference between two years ago and today, he adds, is that agents are now genuinely capable with anything resembling a file system, which opened the door to persistent memory architectures and longer-horizon tasks.

I think the shift is been at the beginning of this year clearly that agentic coding went from being kind of useful to really changing everything. Martin Casado

Geoffrey Irving, a researcher at Anthropic, describes the transition at the capability level rather than the metric level. Full agentic tasks that were not feasible a few weeks before his remarks became feasible almost overnight, he says. Cat Wu, describing code review work, is more specific about what drove the change in her context: it was only with what she calls Opus 4.5 and 4.6, and Sonnet 4.6, that her team felt confident running multiple code review agents simultaneously to traverse an entire codebase and synthesize a set of real issues for an engineer to address before merge. Roman Chernin, reflecting on the broader agentic use-case landscape, notes that coding “started working like maybe few months ago” as the one use case that is genuinely reliable at scale, against a field of many that are not.

The behavioral consequences inside engineering organizations are already visible. Tobi Lütke, Shopify’s chief executive, says that many of the company’s best engineers have not written code this year, noting that December changed everything. Wesley Huff separately reports a secondhand claim attributed to Spotify: that none of Spotify’s engineers have written a line of code since December. That claim is reported rather than confirmed and should be read accordingly. Marc Andreessen, the co-founder of Andreessen Horowitz, puts a productivity figure on the shift: leading-edge programmers are roughly 20 times more productive than they were a year ago. Dylan Patel, a semiconductor analyst, describes Anthropic’s models as having advanced from the capability level of an L4 engineer to an L6 engineer in two months.

Brian Armstrong, Coinbase’s chief executive, offers a ground-level illustration of what that productivity looks like in practice. He describes asking a high-end model to break a feature into three phases, each containing 10 components. He then instructed it to spin up 10 parallel agents to execute phase one simultaneously, mixing in open-source models and Grok to manage cost. About two minutes before he went on air, the system notified him that all 10 components from phase one were complete and ready for review. Armstrong sees this as the early shape of a recursive self-improvement loop in software development, where the acceptance rate for one-shotted pull requests ticks upward over time as agents learn.

What the evidence does not yet settle is how durable the current pace of improvement will be. The shift from 16 percent to 80 percent commit share at Devin, from 1,200 to 7,000 weekly pull requests at Stripe, from L4 to L6 capability at Anthropic’s models in two months: each of those curves is steep enough that extrapolating them even a few quarters forward produces numbers that strain credulity. What remains open is whether the organizations built around human-authored code will reorganize quickly enough to absorb the change, or whether the gap between what agents can now do and what those organizations are structured to use will keep widening on its own.