AI coding agents make code review more important, not less
The productivity gains from AI coding agents are real, but so is the dependency: without human review, output drops toward zero. The evidence from practitioners points in one direction, and it is not toward fewer eyes on the code.
The failure mode Walden Yan describes is precise enough to be useful. Without code review, an AI-assisted codebase becomes unmanageable within roughly two weeks. By the end of that window, a task as simple as changing a button color breaks down because that button has been implemented in ten different places. The codebase, as Yan puts it, “regresses to your worst engineer”: the team member who moves fast without auditing their output has their patterns cemented into the code, and the AI then references those patterns across everything it touches afterward. Yan’s conclusion is direct: “the idea of like you don’t have to look at code I think is generally a bad idea.”
The problem is not confined to individual teams moving carelessly. swyx points to a Meter blog post finding that approximately half of SWE-bench code that passes benchmark tests is completely unmergeable. That is the benchmark problem in its starkest form: a system can clear the tests and still produce output that no engineering team could ship. Passing the test and being fit for production are not the same thing, and the gap between them is exactly what human review exists to close.
At the level of open source maintenance, Gavriel Cohen frames the dynamic as an arms race. Coding agents have made it exponentially easier to open pull requests, and the burden of triaging, reviewing, and verifying alignment has grown faster than the tooling to handle it. Jean-Baptiste Kempf echoes this from the maintainer side, describing the resistance of curl maintainer Daniel Stenberg to what Stenberg calls “AI slop”: fake or low-quality reports and patches that increase the burden on maintainers who are already stretched. The volume of submissions has outpaced the human capacity to evaluate them.
The meme that I have is that your codebase regresses to your worst engineer because that engineer who is, you know, very gung-ho about AI and is not auditing their code, their pattern starts cementing into the code and now the AI is referencing their patterns. Walden Yan
Security adds another dimension. Aaron Levie reports that when AI built roughly 80 to 90 percent of a feature, the thing that slowed the release was not integration or testing. It was a full security review, required because the team could not risk accidental code injection in what had been generated. The AI accelerated the build phase and concentrated the risk into the review phase, making that phase the critical path rather than eliminating it.
The tooling is catching up, though unevenly. Cat Wu reports that running multiple code review agents simultaneously to traverse an entire codebase and surface real issues only became feasible with what she describes as “Opus 45 and 46 and Sonnet 4.6.” Before those models, the reliability was not there for production use. That timeline matters: the coding agents arrived and scaled before the review infrastructure was ready to match them, which is one structural explanation for the quality debt that teams are now carrying.
Nathan Labenz captures the underlying dependency with a figure that reframes the productivity conversation. Practitioners, he reports, estimate roughly a 2x productivity gain from AI. But the same practitioners acknowledge that without human oversight, their output would drop to near zero. The multiplier and the dependency are inseparable. You cannot take the gain without taking on the review obligation that makes the gain real rather than illusory.
Dan Balsam extends this into a structural argument about the labor market. Despite the rise of coding agents, the number of software jobs has increased rather than declined. His explanation is that as models become more capable, steering them effectively requires even higher judgment from the humans working with them. The skill being demanded has shifted upward, not disappeared. Dan Shipper observes a parallel shift in data science: practitioners whose jobs have moved toward reviewing AI-generated work rather than producing original analysis from scratch. The pattern is consistent. The work is not going away. It is changing into something that requires more careful attention, not less, from the people doing it.