Mohith Shrivastava, a principal developer advocate at Salesforce, has a name for what's happening to the humans in this equation. In a Fortune op-ed published last week, he called it the rise of the "supervisor class" — developers whose primary value is no longer writing code but orchestrating the autonomous agents that do. The term is his coinage. The underlying shift is not.
The evidence for it is scattered across enterprise deployments and at least one company that has already moved. At Stripe, internal AI coding agents now autonomously generate and merge more than 1,000 pull requests per week, handling routine engineering work while developers focus on higher-level architectural decisions, CIO reported. Whether most engineers will shift to orchestration soon is unknowable from public data. The direction is not: the job is changing, and at some companies it already has.
The infrastructure enabling this shift is real. Salesforce's Agentforce platform generated 2.4 billion Agentic Work Units to date, with 771 million in Q4 2025 alone — a 57 percent quarter-over-quarter jump. The platform closed 29,000 Agentforce deals in that quarter, up 50 percent from the prior period, and Agentforce ARR reached $800 million. Lennar, the U.S. homebuilder, now runs 1.1 million agentic workflows per month through the platform — not a pilot, not a proof of concept, production volume at scale. Ori Klein, Lennar's VP of digital product development and marketing, described the result simply: they are now available to customers "wherever they are, on their phone, computer, any time of day."
But the infrastructure story has a quality counterpart that vendor announcements tend to skip. The Google DORA 2025 report found that teams with high AI coding adoption — roughly 90 percent — correlated with a 9 percent increase in bug rates, a 91 percent increase in code review time, and a 154 percent increase in pull request size, developer and researcher Mike Mason noted in January, synthesizing the DORA data. Bigger PRs. More reviews. More bugs getting through. The efficiency story is real. So is the quality debt accumulating alongside it.
A December 2025 study by researchers at UC San Diego and Cornell (published as an arXiv preprint) found that professional software developers working with AI coding agents don't go passive. They retain agency, insist on quality attributes like test coverage, and deploy explicit control strategies to manage agent behavior. The supervisor class is not a passive replacement tier. It is a job that requires holding the line.
The metrics used to track this transition are worth scrutinizing. The Agentic Work Unit — Salesforce's core unit of agentic activity — tracks execution, not accuracy. A triggered workflow or an API call counts, regardless of whether the agent resolved the issue correctly. One analyst at Moor Insights & Strategy noted the metric measures activity, not quality. The 2.4 billion AWUs is a volume number. What fraction actually solved the problem they were supposed to solve is a different question that the number doesn't answer. Salesforce's own internal operations show a 96 percent autonomous resolution rate; the company's homepage says 66 percent across customer deployments broadly — two different populations, two different numbers.
The structural shift the supervisor class thesis describes is real. The evidence for it is in production deployments and the DORA data. What the evidence doesn't yet show is that the transition is clean, or that the humans being restructured are the only ones who need to adapt. The code being written by agents is getting bigger, reviewed more slowly, and shipping with more bugs. The supervisor class has a job description that isn't fully written yet.
