Nicholas Arcolano

Nicholas Arcolano is Head of AI & Research at Jellyfish, where he measures how coding assistants and autonomous agents change software development across hundreds of companies. His research confronts a persistent management problem: AI can accelerate coding without delivering equivalent improvements in completed projects, software quality, or business outcomes.

Arcolano earned a doctorate at Harvard, where he coauthored research on estimating covariance-matrix principal components. He worked at MIT Lincoln Laboratory, Runkeeper, and TrueMotion before joining Jellyfish, initially in data science. By 2022, he led research spanning capacity planning, delivery prediction, machine learning, and engineering analytics.

As generative AI reshaped software development, his team combined coding-tool activity with source-control and project-management data to study adoption in working organizations. Its datasets expanded from more than two million pull requests in 2025 to more than 37 million by April 2026, revealing widening differences between companies adopting autonomous coding agents and those still experimenting.

What his research reveals

• Behavior-based AI adoption: Arcolano measures how often developers actually use AI while coding, avoiding misleading proxies such as licenses purchased or generated lines. In an analysis of 20 million pull requests, median adoption rose from approximately 22% to nearly 90%, while autonomous agents produced fewer than 0.2% of merged pull requests during the period studied.

• Repository architecture shapes productivity: His research on distributed codebases uses active repositories per engineer to compare organizational complexity independently of team size. Centralized architectures showed stronger productivity gains than highly distributed systems, where agents struggle to assemble context across repositories, dependencies, and undocumented service boundaries.

• Engineering outcomes over activity metrics: Full adoption correlated with approximately twice the pull-request throughput, but downstream project delivery increased by an average of just 27%. Reviews, testing, planning, and cross-team coordination can consume the apparent coding gains. AI-assisted pull requests were also approximately 18% larger, without a statistically significant association with new bug tickets or reversions in the observed dataset.

• The tyranny of the token curve: Arcolano challenges the assumption that consuming more model tokens reliably increases developer output. His analysis of seniority and token spending found diminishing returns for junior and senior developers alike; higher spending widened their productivity gap instead of helping less-experienced engineers catch up.

His practical focus is organizational readiness: giving agents usable context, managing review bottlenecks and inference costs, and determining whether faster coding produces work that actually reaches customers.