AI Writes the Code. So What's Left for Engineers? - Laurie Voss, Arize AI

Summary

Laurie Voss argues that AI coding agents are making code generation dramatically cheaper, but they do not eliminate the work of software development. Instead, they shift the bottleneck to tasks that remain difficult to automate.

How the bottleneck moves

• Voss cites figures suggesting that developers using autonomous coding agents produce 741% more code, while products ship only 30% faster. He attributes much of the gap to the time and effort required to verify generated code.
• Human review does not scale easily. He refers to a study finding that reviewers’ ability to understand code drops sharply after about 450 lines per hour.
• AI code review is already in use, but Voss says current reviewers find different issues and are not reliable enough to trust alone. He points to reports of failures after teams skipped review and to Amazon’s requirement that senior engineers sign off on review results.
• Passing tests is not the same as producing a change that maintainers consider safe to merge. Voss cites a benchmark where success fell from 87% on tests to 54% on whether a reviewer would merge the change. Merge decisions depend on context such as code style, side effects, external systems, and repository history.

Drawing on the theory of constraints, Voss says that solving one bottleneck will expose another. After code review, he expects operations—deployment and scaling—and then security to become major challenges. He predicts progress on code review in the near term, while operations, scaling, and security may take much longer to automate.

What remains human-centered

Voss’s central claim is that durable work involves discovering what people actually need, defining what a good solution means, and shaping software into a usable product. A request such as “track my orders” is too vague: the right solution depends on the user’s business and its processes. That knowledge is often held by the people doing the work, so engineers must learn it through direct engagement.

He connects this to essential complexity: tools can reduce incidental work such as typing and syntax, but they cannot remove the complexity inherent in understanding and solving the problem. He cites Fred Brooks’s ideas about software complexity and Sean Grove’s view that code is a lossy projection of a specification.

Some review knowledge can be made explicit and fed back into automation. For example, Voss says Cursor BugBot turns team feedback into rules; he cites a collection of roughly 44,000 rules. But he argues that the broader task of defining requirements and quality remains crucial.

The “product engineer” role

Voss calls the emerging role the product engineer: someone who understands users, decides what should be built, defines quality, and uses agents to implement it. He sees this as a blending of engineering and product work, rather than simply a new kind of AI specialist.

He cites a reported 800% increase in customer-facing engineering job openings in the first nine months of 2025 and an average salary of about $240,000, while noting that job titles for similar roles vary. In his view, software development and programming are becoming distinct: agents may do more of the programming, while engineers remain responsible for product decisions and outcomes.

The junior developer challenge

Voss warns that traditional junior training—assigning small coding tasks and reviewing the results—is being displaced by agents. He points to labor-market data showing declining employment among engineers aged 22–25 and rising demand for more experienced engineers.

He argues that the industry needs to deliberately train people to understand users, elicit requirements, and write specifications—not just teach programming languages or frameworks. He cites Google’s Associate Product Manager program as a relevant model, but says its scale is far below the number of people he believes will be needed. He also acknowledges that developers who love the craft of coding may dislike a job focused more on customer discovery and product definition.

Practical recommendations

For engineering teams, Voss recommends:

1. Prepare for less human involvement in routine code creation. Focus human review on critical code and use automated reviewers for the rest.
2. Automate whatever can be verified cheaply. Build automated ways to assess quality rather than relying only on manual processes.
3. Turn expert judgment into explicit specifications, criteria, and evaluations that agents can use.
4. Measure outcomes, not code volume. Track whether review comments are addressed, how often passing tests leads to a merge, and what happens in production.
5. Treat production behavior as the ultimate test. Monitor deployed software and evaluate its real-world performance.

Voss’s overall forecast is that the next decade will bring significant disruption, but software development will likely expand into a larger, product-oriented profession as its definition changes.

Main speaker and sources referenced

• Main speaker: Laurie Voss, head of developer relations at Arize AI and co-founder of npm.
• Sources and people referenced: Eliyahu Goldratt and the theory of constraints; Fred Brooks; Sean Grove of OpenAI; Cognition’s code-review benchmark; GitHub Copilot; Cursor BugBot; Amazon; Anthropic; Andrew Ng; Gergo Oros; and Dexter Horthy.