AI coding agents are creating fragile codebases and more outages
A former Meta developer warns that relying on AI for code generation leads to shallow tests, spaghetti code, and longer production outages.
Tools ยท Source: Hacker News
What happened
Evan X. Merz, a developer recently laid off from Meta, claims the industry push toward AI generated code is creating brittle software. Merz says he generated 90 percent of his code using AI at Meta just to meet performance expectations. He notes that this massive output did not actually improve his productivity by ten times or even two times. Now at a new company, he generates 100 percent of his code with AI because that is the baseline expectation today. He warns this mandate is breaking large scale SaaS products. Merz watched top programmers merge terrible code, like using JavaScript events for simple CSS hover effects.
Merz identifies four specific ways AI breaks codebases. First, AI writes shallow tests that only cover trivial methods or surface level interactions. Second, it writes overly complex code that follows rules to a fault, adding unnecessary helper methods. Third, humans lose their understanding of the system as AI generates more of the logic. Fourth, AI moves too fast. Pull requests arrive faster than human developers can review them. Developers skim the code and miss crucial details, allowing technical debt to pile up into spaghetti code.
The results are measurable in production environments. Merz notes that AI heavy teams see a slow but steady rise in critical incidents. At first, the pull requests merge smoothly. Then the priority zero bugs start rolling in. When these outages happen, recovery takes much longer. Developers have to dig through complex, unfamiliar code they did not write. Errors slip past automated tests and are only discovered by frustrated end users. The final result is a degraded user experience built by machines for machines.
Key facts
- 90% โ The minimum amount of code Merz generated with AI to get a decent rating at Meta
- 100% โ The amount of code Merz generates with AI at his new tech company
- May โ The month Merz was laid off from Meta
Why it matters
We are trading long term stability for short term velocity. AI coding agents excel at solving problems that fit neatly inside a small context window. They fail completely at understanding system wide architecture or long term maintainability. When developers use AI to write all their code, they stop understanding the underlying data structures. Reviewing pull requests becomes a dangerous rubber stamp process. The codebase turns into a black box that nobody on the engineering team fully comprehends. You cannot build a resilient SaaS application if your team does not know how the pieces fit together.
The second order effect is a massive increase in maintenance costs and production downtime. Companies think they are saving money by shipping features faster with AI agents. Instead, they will pay a heavy tax on incident response. When a critical bug hits production, the team will spend hours just trying to understand the AI generated spaghetti code. Recovery times will skyrocket. Eventually, high quality software will become a premium product defined by human craftsmanship. The cheapest products will rely entirely on AI and suffer from constant outages, while premium products will win by offering a stable, human tailored experience.
For builders
Audit your automated test coverage
Do not trust AI to write meaningful automated tests. AI agents default to testing trivial getters and setters while ignoring deep business logic and edge cases. Founders pay the ultimate price when unhandled exceptions crash the product in production. You must mandate human review for all test coverage.
Enforce strict pull request reviews
AI generates code much faster than humans can read it. You must allocate dedicated time for developers to actually understand the code they are merging. Teams that skim pull requests will drown in technical debt and face constant production outages.
Watch your incident recovery times
Track exactly how long it takes your engineering team to resolve critical bugs. If your mean time to recovery is climbing month over month, your team no longer understands your codebase. You will lose enterprise customers to competitors with more stable products.
My take
I see this exact scenario playing out in every startup right now. Founders are completely addicted to the speed of AI coding tools, but they are blind to the rot accumulating under the hood. If your engineers do not understand the code they are shipping, you do not have a product. You have a ticking time bomb.
Original reporting: Hacker News. This is my rewrite and opinion.