Chelsea Troy argues that the tech industry measures AI productivity through the wrong lens. Rather than focusing on code velocity, teams should leverage AI agents to conduct the engineering work that time constraints previously made impossible.

Troy, a Mozilla MLOps engineer and University of Chicago instructor, contends that agentic AI systems unlock value not by accelerating shipping timelines but by enabling rigorous experimentation, testing, and simulation. Engineering teams have long wanted to run comprehensive validation workflows, stress tests, and exploratory analyses. Insufficient bandwidth consistently forced tradeoffs in favor of shipping code over quality assurance.

Agentic systems change this equation. These AI agents operate autonomously within defined parameters, handling repetitive experimental work that humans would normally coordinate manually. An agent can spin up multiple test scenarios, analyze edge cases, run performance benchmarks, and generate detailed reports without human intervention between steps.

The distinction matters for how organizations should allocate AI labor. Current productivity frameworks focus on lines of code per engineer or features deployed per sprint. These metrics miss the core benefit: teams gain capacity to answer questions they previously couldn't afford to investigate. Why does this service degrade under specific load conditions? How would the system behave if we modified this algorithm? What happens when we test this edge case across 1,000 variations?

Troy's framework reframes AI's workplace impact. Rather than replacing developers or accelerating completion of planned work, agentic AI systems function as force multipliers for research and validation. They handle the computational grunt work, freeing engineers to design better experiments and interpret results.

This perspective carries implications for hiring, training, and ROI measurement. Companies investing in agentic AI should expect longer-term value from improved system reliability and faster feedback loops rather than immediate sprint completion gains. Engineering teams need skills in experiment design and result interpretation, not just code generation.

The real productivity windfall arrives when teams move beyond asking "How fast