AI coding agents are making codebases worse, and no amount of tooling can fix it

AI Coding Agents Degrade Codebases, and Tooling Can't Fix It

The AI coding industry has a consistent prescription for agents that produce messy code: add more guardrails. Stack up the linters, deploy the review bots, configure the autonomous loops. The promise is that enough scaffolding around the model delivers speed without the mess.

The data says you get the mess anyway.

Faros AI, an engineering productivity and intelligence company, published its 2026 AI Engineering Report built on two years of telemetry from 22,000 developers across 4,000 teams. The findings describe an industry shipping more code than ever while quality erodes across every metric they tracked. Bugs per pull request rose 28 percent. Incidents per PR tripled. Median review times multiplied fivefold. Code churn hit 10x. On their research page, Faros frames the pattern in one line: "Volume is up, quality is down, and the gap between the two is widening as adoption deepens" 1.

The report says its methodology relies on telemetry rather than self-reported surveys, comparing outcomes at low versus high AI adoption within the same organizations. It also claims its data "directly contradicts DORA's 2025 findings," the widely cited DevOps Research and Assessment benchmarks from Google, and that even senior engineers are not closing the quality gap 1. But Faros AI sells engineering productivity tooling. The study is vendor research, not peer-reviewed. And the person circulating it most aggressively concedes the point.

That person is Dex, founder of HumanLayer, a company building tools in the human/agent collaboration space. In an essay titled "Why Software Factories Fail," he cites the Faros data and then makes a bigger claim: the problem is not that developers need better tooling around AI coding agents. The problem is in how the models themselves are trained and evaluated. His central argument: "no amount of harness engineering or loopsmaxxing can solve what is fundamentally a model-training issue" 2. Harness engineering, as he uses the term, covers the full apparatus of linters, review bots, and autonomous coding loops that companies assemble to catch agent mistakes. The claim is that these guardrails treat symptoms while the cause sits in the training pipeline.

Dex discloses his competitive interest at the top of the essay. HumanLayer sells products for keeping humans involved in the agent workflow, and a world where harness engineering alone fixes AI coding would be a world that does not need his company 2. He also concedes that the Faros report is "more of a correlation signal than a verifiable smoking gun" 2. The report can show that AI tool adoption correlates with declining quality metrics. It cannot prove that the tools cause the decline.

Even with those caveats, the circumstantial case keeps growing. The essay points to StrongDM's "lights-off software factory," a system launched with the stated narrative that humans are the bottleneck, the models are good enough, and code is essentially free 2. Dex writes that he has been unable to find any definitive results from the project. The factory's weather report shows sparse updates between February and June. A July conversation on Hacker News with the team suggests a more formal update may be coming 2.

A company that built its narrative around autonomous code generation has gone months without publishing outcomes. OpenAI is heading down the same path. Ryan Lopopolo of OpenAI published an essay on harness engineering in February and gave a talk in April about the company's internal software factory, called Symphony 2.

The defense circulating in AI coding communities is that poor agent results are a skill issue. If your agents produce bad code, the argument goes, you need to prompt better or spend more tokens. Dex rejects this, admitting he used to make the argument himself before changing his mind 2. Other practitioners are landing in the same place independently. At AI Engineer Europe, a developer who goes by Mario warned the audience to slow down, citing companies experiencing outages from coding-agent mishaps. Matt Pocock has said that codebases are deteriorating faster than at any point before 2.

The so-what for builders, investors, and engineering leaders: if the quality problem originates in model training rather than tooling configuration, then no linter or review bot resolves it. The fix has to come from the labs, in how coding models are trained, evaluated, and benchmarked. Until then, every autonomous coding pipeline is accumulating a specific kind of debt: code that ships fast, breaks in production, and resists the guardrails designed to catch it.

The Faros data, correlation or not, suggests that debt is already compounding. If the root cause is in model training, the most important fix is the one no amount of guardrails can deliver.