The bottleneck is not code. It never was.

The core claim. Laurie Voss argues that software engineering’s center of gravity has shifted permanently. Writing code is effectively free, thanks to AI agents. What remains expensive, and what humans must own, is everything before and after the typing—understanding the problem, defining what “good” means for a specific context, and making software that people actually enjoy using. This is not a prediction about a distant future; Voss argues the transition is already well underway.

The data that matters. Voss cites numbers that cut through the hype. Developers using autonomous coding agents write 741% more code, but ship only 30% faster. The delta is not a failure of the tools—it is a classic theory-of-constraints problem. The bottleneck has moved from writing code to reviewing it, and it will keep moving. Voss draws on a 20-year-old Cisco study: human code review accuracy collapses after about 450 lines per hour. When agents produce 10,000-line diffs, review becomes impossible.

What is being tested today. Automated code review is already in production at major companies. GitHub Copilot handles more than one in five pull request reviews. OpenAI says humans may review PRs but are not required to. Yet the evidence is mixed. When researchers benchmarked four AI reviewers, each found real bugs—but no two found the same one. Cognition’s Devin benchmark shifts the question from “does it pass tests?” to “would a maintainer merge this?” and pass rates drop from 87% to 54%.

The bet on what agents cannot do. Voss’s provocative thesis is that agents will eventually handle code review, operations, scaling, and security, because those activities can be captured in data and trained against. The bottleneck that will not move, he argues, is discovering what people actually need and defining what “good” means in a particular business context. That knowledge exists only in the heads of domain experts and customers. It cannot be extracted from training data.

The organizational shock. This shift is not neutral. The job of “product engineer”—someone who talks to customers, writes specs, and ships with agents doing the typing—is emerging across the industry. Job postings for forward-deployed engineers, a closely related role, rose 800 percent over nine months in 2025. But the transition is already creating casualties. Employment data from January 2023 shows engineers aged 22-25 in a steep decline. The structured tickets that used to train junior developers have been given to agents.