The promise of artificial intelligence transforming software development is not merely hype; it is a measurable reality, albeit one with nuanced dependencies. Nicholas Arcolano, Head of Research at Jellyfish, presented compelling data insights from an extensive dataset comprising 20 million pull requests from 200,000 developers across 1,000 companies, spanning from June 2024 to the present. His analysis reveals that while AI coding tools are rapidly integrating into workflows, the extent of productivity gains is profoundly shaped by an often-overlooked factor: the underlying code architecture.
Across the software industry, the adoption of AI coding tools has surged dramatically. Starting from a modest 2% in June 2024, the proportion of companies generating 50% or more of their code with AI has ballooned to nearly half by October 2025. This rapid embrace extends to individual developers, with median AI adoption rates skyrocketing from 22% to approximately 90% within the same timeframe, indicating a widespread shift in coding practices.
However, not all AI tools are created equal in their current impact. Arcolano highlighted that fully autonomous coding agents remain in their infancy. "Only 44% of companies have used a fully autonomous agent in the past 3 months," he stated, emphasizing that this usage is predominantly for trialing and experimentation, not full-scale production, accounting for less than 0.2% of all pull requests. The real, immediate gains are coming from interactive AI coding tools.
These interactive tools are delivering substantial productivity enhancements. Companies leveraging AI are experiencing a remarkable 2x increase in pull request (PR) throughput and a 24% reduction in cycle times. Arcolano noted, "The average trend here is about a 2x change as you go from zero to full adoption," confirming that doubling PR output is a realistic expectation for widespread AI integration. This acceleration means more code is being shipped faster, a critical metric for any agile development team.
Beyond speed, the nature of the code itself is evolving. Data shows an 18% increase in the average size of pull requests, specifically in net lines of code added, rather than merely reworked or deleted lines. Interestingly, this surge in output has not, to date, demonstrably compromised code quality. Jellyfish’s research found no statistically significant relationship between the rate of AI adoption and an increase in bug tickets created or PR reverts. In fact, teams are increasingly using AI to tackle and resolve existing bug backlogs, suggesting a positive, if indirect, impact on code health.
The most critical insight, however, pertains to how a company’s code architecture dictates the magnitude of these AI-driven productivity gains. Arcolano categorized architectures into four regimes based on "active repos per engineer": centralized, balanced, distributed, and highly distributed. When analyzing PR throughput across these categories, a stark divergence emerges. "What's really interesting is this highly distributed case. There's essentially no correlation here between AI adoption and PR throughput," Arcolano revealed. Teams with centralized or balanced code architectures saw productivity gains of up to 4x, far exceeding the 2x average. Conversely, highly distributed architectures showed negligible, or even slightly negative, correlation between AI adoption and PR throughput.
Related Reading
- Architecture Copilots: The Overlooked Frontier for Enterprise ROI
- Claude Code Redefines Developer Workflows on Desktop
- The Unseen Architect: How AI Can Engineer a Future of Zero Bugs
This disparity stems from the "context problem." Current interactive AI coding tools are optimized to operate within a single repository, making it challenging for them to maintain context across multiple, disparate repositories. In highly distributed architectures, where engineers frequently work across numerous repos, the AI’s ability to provide relevant, high-quality suggestions diminishes, thereby negating its productivity benefits. This suggests that the structural organization of a codebase can either amplify or mute the transformative power of AI.
For founders, VCs, and AI professionals, these findings underscore that successful AI transformation requires more than just deploying tools. It demands a strategic alignment of AI adoption with existing code architecture, or a deliberate effort to optimize architecture to harness AI’s full potential. Without addressing the foundational structure of their codebases, organizations risk investing heavily in AI tools without realizing their promised productivity dividends.
