The long-standing management principle, “What you measure matters,” is being intensely re-examined within software engineering, particularly with the advent of AI coding agents. For decades, engineers have debated productivity metrics, often focusing on quantifiable outputs like lines of code. However, as the new generation of AI coding agents delivers unprecedented volumes of code, the clarity around effective measurement has diminished.

One emerging, yet problematic, metric is the “token budget”—the authorized amount of AI processing power a developer can consume. While these large budgets have become a status symbol among Silicon Valley developers, measuring an input rather than an output is an unconventional approach to productivity. Such a metric might encourage AI adoption or token sales, but it fails to accurately reflect efficiency or value creation.

Evidence from a new class of companies specializing in “developer productivity insight” highlights this growing disconnect. Firms analyzing tools like Claude Code, Cursor, and Codex observe that developers using these AI agents generate significantly more accepted code than before. Crucially, however, they also find that engineers are forced to revise this accepted code much more frequently, thereby undermining the initial claims of increased productivity.

Alex Circei, CEO and founder of Waydev, a company that provides developer analytics and works with over 50 customers employing more than 10,000 software engineers, has built an intelligence layer to track these dynamics. He notes that while engineering managers often see initial code acceptance rates between 80% to 90% for AI-generated code, they often overlook the significant churn that occurs in subsequent weeks when engineers have to revise that code. This revision process dramatically reduces the real-world acceptance rate to a mere 10% to 30% of the originally generated code.

Founded in 2017, Waydev has completely revamped its platform in the last six months to address the rapid proliferation of AI coding tools. The company is now rolling out new tools designed to track the metadata generated by AI agents, offering analytics on code quality and cost. This provides engineering managers with deeper insights into both the adoption and true efficacy of AI in their workflows. While analytics companies have a natural incentive to highlight existing problems, the mounting evidence suggests that large organizations are still struggling to utilize AI tools efficiently. Major industry players are taking notice; Atlassian, for instance, acquired engineering intelligence startup DX for $1 billion last year to help its customers understand the return on investment of coding agents.

Data collected across the industry paints a consistent picture: more code is being written, but a disproportionate amount of it fails to endure. GitClear, another company in this space, reported in January that while AI tools boost productivity, their data revealed that “regular AI users averaged 9.4x higher code churn than their non-AI counterparts”—a figure that more than doubles the perceived productivity gains. Similarly, Faros AI, an engineering analytics platform, analyzed two years of customer data for its March 2026 report, concluding that code churn (lines of code deleted versus lines added) had increased by a staggering 861% under conditions of high AI adoption.

Jellyfish, an intelligence platform for AI-integrated engineering, collected data from 7,548 engineers in the first quarter of 2026. Their findings indicated that engineers with the largest token budgets produced the most pull requests, but this productivity improvement did not scale efficiently. These engineers achieved only two times the throughput at ten times the cost in tokens, suggesting that the tools are primarily generating volume rather than genuine value.

These statistics resonate with developers, who, despite reveling in the freedom offered by new AI tools, are finding themselves grappling with increasing code review burdens and accumulating technical debt. A notable difference has been observed between senior and junior engineers, with the latter tending to accept significantly more AI-generated code, consequently leading to more extensive rewriting later on. Nevertheless, even as developers strive to fully understand the behavior and output of their AI agents, there is no anticipation of a reversal. As Circei stated, “This is a new era of software development, and you have to adapt, and you are forced to adapt as a company. It’s not like it will be a cycle that will pass.”