Chelsea Troy, a Mozilla MLOps engineer and a University of Chicago instructor, argues that the tech industry is measuring AI-driven developer productivity the wrong way. In a conversation on Ben Lorica’s Generative AI in the Real World podcast, Troy says the main opportunity of agentic coding is not simply shipping more code faster, but having the bandwidth to run the experiments, tests and simulations teams have long wanted but lacked time for.
Who uses agentic tools and why it matters
Troy balances three roles: she helps machine learning engineering teams get models into production at Mozilla, she teaches courses at the University of Chicago, and she develops content for O’Reilly about agentic coding skills. In all three settings she finds it important to understand how agentic coding tools work, because colleagues and students use them and she needs to advise on their integration into workflows.
At Mozilla they are exploring how well agentic coding fits organizational values and internal workflows. Troy’s most nuanced perspectives have come from working with graduate computer science students who must navigate entry-level roles, applications, evaluations and early promotions in an environment where these tools are commonly available.
Teaching Python: focus on why, not just syntax
Troy notes substantial development in programming-focused agentic tools between 2025 and 2026. As a result she redesigned her Python and applied data analysis courses to emphasize the reasons behind Python’s design choices and when Python is appropriate, rather than focusing primarily on syntax memorization.
She argues Python is a valuable case study because of its relatively small token set, strong syntactic patterns and extensive public documentation about interpreter changes and design trade-offs. That public record makes Python a useful example for teaching how to grow and maintain a large, complex codebase.
Interviews and inherited questions
Troy explains that many common programming interview questions—such as implementing a linked list from scratch—originated as pragmatic tasks when higher-level language support did not exist. Over time the original practical rationale was forgotten and the questions remained, sometimes justified retroactively as tests of critical thinking.
As an interview designer for senior roles at Mozilla and as an advisor to students, Troy sees the landscape from two angles. For senior and staff positions she uses live interviews that emphasize architectural trade-offs and decision-making rather than lines of hand-written code. She believes decision-focused screens are more aligned with on-the-job requirements in the present era.
Entry-level hiring: uncertainty and experimentation
Troy says it is currently challenging for students seeking their first industry jobs. Some companies are experimenting with whether they still need entry-level hires if senior engineers can delegate to agentic tools. While Troy expects firms will ultimately continue to invest in junior talent, the current experimental period has increased anxiety among early-career applicants.
Managers coding again and career flexibility
Some managers are excited that agentic tools let them re-engage with individual contribution without returning to full-time hands-on coding. Troy sees this as an opportunity to rethink career paths—potentially enabling more movement between individual contributor work and leadership roles rather than the traditional one-way transition to management.
Quality over volume: running experiments instead of guessing
Troy is critical of productivity narratives that equate success with shipping more tickets, more code, or a higher rate of closed items. In her experience teams often make architectural choices without running exhaustive tests or comparisons because they lack bandwidth. Agentic tools could enable teams to try multiple alternatives, run simulations and select solutions based on outcomes and data rather than educated guesses.
Data visualization as a practical example
Troy has been experimenting with agentic tools to generate HTML/CSS visualizations for teaching and dashboards. Lacking deep frontend fluency, she uses language models to produce baseline visualizations and then iteratively refines them, which reduces development time and increases the number of designs she can test—an example of how these tools expand experimental capacity.
Shifting metrics: from tokens consumed to token efficiency
Looking ahead two years, Troy hopes the industry’s measurement focus will shift away from raw token consumption as a productivity metric toward token efficiency as a measure of skill in using agentic tools. She expects token-based leaderboards to become less attractive as usage costs rise, and hopes organizations will instead reward effective, outcome-driven tool use.
Conclusion: tooling that enables experimentation
In sum, Troy sees agentic coding as a means to restore bandwidth for experimentation and rigorous comparison of alternatives—leading to better-engineered outcomes. She also highlights transitional challenges: uncertainty for entry-level hiring, the need to update interview practices, and the importance of investing in platforms and processes that let teams run and measure many experiments rather than relying on single-guess engineering decisions.
This article summarizes the discussion between Ben Lorica and Chelsea Troy on the Generative AI in the Real World podcast, which examines changes in education, hiring and productivity metrics as agentic coding tools spread.
