AI Coding Tools Not Paying Off? Your Code Architecture Might Be to Blame (jellyfish.co)

Jellyfish (with OpenAI) analyzed 321 customers from Jan–Aug 2025—over 3.8 million PRs across 130,000 repos and 9,602 company-weeks—to measure how AI coding tool adoption affects developer throughput. They define AI adoption as the fraction of active coding days an engineer used AI and introduce “weekly active repositories per engineer” to characterize code architecture (quartiles: Centralized <1, Balanced 1–2, Distributed 2–3.7, Highly Distributed 3.7+). Broadly, full AI adoption corresponded to a 2.1x increase in PRs merged, but effects vary sharply by architecture: centralized and balanced teams see ~4x change going from 0→100% adoption, distributed teams ~2x, while highly distributed teams show noisy or slightly negative correlations. Independent of AI, highly distributed orgs merge ~5.3 PRs/engineer versus ~2.7 in centralized orgs, largely due to cross-repo coordination overhead. The authors argue the gap comes from context availability: LLM-driven tools perform better when code and dependencies are consolidated and easily indexed, whereas polyrepo/microservice landscapes demand complex, cross-repo coordination that current tools struggle to automate. Practical implications: teams expecting big AI returns should evaluate repo topology and invest in context engineering (indexing, search), cross-repo agents, and validation pipelines; highly distributed organizations may need bespoke automation and agent workflows to realize comparable gains.