Chelsea Troy, a Mozilla MLOps engineer and University of Chicago instructor, challenges how the tech industry measures AI productivity gains. Rather than focusing on code velocity, Troy argues the real value lies in what agentic AI systems enable teams to do with newly freed capacity.

The conventional narrative celebrates AI tools that generate code faster. Developers ship features quicker. Deployment cycles shrink. But Troy reframes the opportunity entirely. When AI handles routine coding tasks, engineering teams gain time to execute work they previously deprioritized: comprehensive testing, rigorous experimentation, and detailed simulations. These activities directly improve code quality and system reliability.

This distinction matters because it separates marketing from actual engineering value. Many organizations measure AI's impact by lines of code written or hours saved. Troy's perspective suggests those metrics miss the point. A team that uses AI to ship code 20% faster but skips testing gains nothing long term. A team that uses AI to automate boilerplate work and redirects that time toward experimental validation and simulation gets compounding returns through more robust systems.

Agentic coding, where AI systems operate with greater autonomy in development workflows, amplifies this effect. These agents can run multiple test scenarios, flag edge cases, and validate assumptions without constant human intervention. The bandwidth savings compound because agents work continuously while humans focus on higher-level architecture and strategy.

Troy's framing reveals a productivity blind spot across the industry. Engineering leaders optimizing for output velocity may actually be degrading system quality. Teams building critical infrastructure cannot afford to trade testing for speed. The sustainable play involves using agentic AI to do both more code and better code, enabled by the bandwidth AI frees up for work that prevents disasters rather than just ships features faster.

Her argument lands during a moment when AI vendors push metrics-heavy narratives around developer productivity. Troy's perspective offers a corrective: real productivity gains in software engineering mean enabling