Laurie Voss, head of developer relations at Arize AI and npm co-founder, examines engineering work after coding agents in this 27-minute AI Engineer talk.

Faster code moves the queue

  • Voss cites developers producing 741% more code while shipping only 30% faster. Code volume and delivered value measure different things.
  • Review becomes the immediate constraint when generated changes outpace people’s ability to inspect them.
  • Automated reviewers can find complementary bugs, but their disagreements and missed defects still require judgment.
  • Following Eliyahu Goldratt’s theory of constraints, he expects each improvement to expose the next bottleneck.

Decide what good means

  • Voss predicts review, operations and security will become increasingly automated over five to ten years.
  • His longer-term bet is that discovering user needs and specifying success will remain the human constraint.
  • “Track my orders” means different things for a bakery, an auto-parts shop and a manufacturer. Learn the business before defining the system.
  • Turn experienced engineers’ judgment into specifications, scoring rubrics and evals, tests of how a model or system behaves.

Train the product engineer deliberately

  • His product engineer discovers needs, defines acceptable outcomes and ships with agents doing much of the implementation.
  • Junior tickets used to provide practice and feedback. Delegating those tasks to agents leaves an apprenticeship gap.
  • Give early-career engineers supervised customer discovery and specification work, alongside the technical foundations needed to judge results.
  • He acknowledges a real loss for people who entered programming for its craft, even while predicting a larger profession centered on product judgment.

Apply the argument with controls

  • Identify high-impact changes that warrant human review; validate automated reviewers before widening their authority.
  • Automate cheap, reliable checks and measure delivery, acted-on review comments and whether maintainers accept the result.
  • Use production monitoring to find failures that escaped earlier tests and turn them into new evaluation cases.
  • Voss’s immediate delegation advice is more aggressive than his evidence about today’s reviewers supports. Keep required approvals and risk-sensitive checks.

Caveat: the workforce outlook is Voss’s forecast. His cited output, hiring and review figures do not establish that AI alone caused labor-market changes or guarantee more engineering jobs.

“Finding out what people need, writing down what good means, and making it something people like.” — Laurie Voss, 26:48–26:51