{"slug": "the-code-is-piling-up-nobody-budgeted-for-the-cleanup", "title": "The Code Is Piling Up. Nobody Budgeted for the Cleanup.", "summary": "New quantitative benchmarks show that agent-generated code is roughly twice as structurally sloppy as human-authored repositories, with verbosity scores averaging 0.33 versus 0.15 and erosion scores of 0.68 versus 0.31. State-of-the-art models tested on iterative coding tasks with context resets achieved a 0% pass rate on strict solve requirements, suggesting AI agents accumulate technical debt across sessions. The findings highlight a tooling gap, as standard CI pipelines lack production-grade checks for verbosity or erosion.", "body_md": "Shipping code has never been faster. Maintaining the codebase that results from that speed is a separate problem that the tools haven't solved — and the data is starting to show it.\n\n## 1. We Now Have Metrics for How Sloppy AI Code Actually Is — And They're Not Pretty\n\nFor a while, the debate over AI code quality was mostly anecdotal. Developers complained about verbosity and pointless abstractions; defenders said it was fine with prompting. Now there's something harder to argue with: [quantitative benchmarks](https://earendil.com/posts/measuring-code-sloppiness/) measuring the structural sloppiness of agent-generated code versus human-authored repositories.\n\nTwo metrics do most of the work here. *Verbosity* measures duplicate and unnecessarily redundant lines. Established human-authored repositories average around 0.15; agent-generated code averages 0.33 — more than twice as verbose. *Erosion* measures how complexity gets concentrated into bloated functions rather than distributed sensibly. Human repos average 0.31; agent code averages 0.68. Again, roughly twice as bad.\n\nThe benchmark also tested state-of-the-art models on iterative coding tasks — the kind where context resets between sessions, mimicking real usage. They achieved a 0% pass rate on strict solve requirements. Not poor performance. Zero. The implication is that AI agents, left to their own devices over multiple sessions, accumulate technical debt at a rate that makes the codebase progressively harder to work with. The code passes tests. It just doesn't stay clean.\n\n**Why it matters:**\n\n**For ICs:** Verbosity and erosion scores give you a vocabulary to push back on agent output — not just \"this feels off,\" but \"this function is carrying 3x the complexity it should.\"\n\n**For leaders:** If your team is using coding agents without a quality gate beyond \"does it pass CI,\" the technical debt is accumulating invisibly. These metrics are a starting point for measurement.\n\n**For founders:** The question isn't whether AI can ship features — it clearly can. The question is whether your team has the maintenance capacity to own what it produces over 12 to 24 months.\n\nThe tooling gap: we have linters, formatters, and test coverage thresholds. We don't yet have production-grade verbosity or erosion checks built into standard CI pipelines.\n\n## 2. Adding Code Is Now Easy. Removing It Has Always Been Hard. That Gap Is Widening.\n\nThere's an older engineering principle that software tends to accumulate features it can't shed — that deletion is harder than addition, and teams almost never do it deliberately. [A recent piece](https://vickiboykis.com/2026/09/01/bad-code-is-kudzu/) uses an ecological metaphor worth stealing: bad code is kudzu. Once it takes root, it spreads. Other code grows around it. Eventually it's load-bearing whether you wanted it to be or not.\n\nThe AI era didn't invent this problem, but it's supercharging the supply side. When generating a new feature takes minutes instead of days, the pressure to prune never catches up. The author made the point with a concrete example: she shipped an ML-powered feature to production, watched it go analytically dark for months, and then pulled it out. Three commits. Surprisingly easy — because she caught it while it was still young. The lesson isn't that deletion is easy. It's that deletion gets harder with every day you wait, and AI is making everyone wait less on the writing side while doing nothing for the reading and removing side.\n\nThe deeper issue is incentive structure. Engineers get credit for features. They rarely get credit for removing them. AI tools optimize for output — tokens generated, tasks completed, PRs open. Nothing in the current toolchain tells you when a feature has gone dark or when a function has grown to the point where it will require a week to untangle. That feedback loop still requires a human paying attention.\n\n**Why it matters:**\n\n**For ICs:** Treat AI-generated code as temporary until proven otherwise. Default to review and prune cycles, not just addition.\n\n**For leaders:** Deletion needs to be a first-class engineering activity — tracked, celebrated, and part of your definition of done.\n\n**For founders:** If you're building fast with AI and not building in a regular pruning discipline, you're accumulating a cleanup bill that compounds.\n\n## 3. AI Sped Up the Part of Software Delivery That Wasn't the Bottleneck\n\n[GitLab's 2026 AI Accountability Report](https://www.infoq.com/news/2026/06/ai-coding-outpaces-governance/) puts a number on something many teams have felt but couldn't articulate: 78% of developers say AI has made them faster at writing code. 79% say overall software delivery hasn't kept pace. Those two numbers belong in the same sentence.\n\nThe mechanism is straightforward: AI tools are exceptional at the text-editor layer. They're irrelevant to the organizational layer — planning, review, testing, deployment, and the overhead of keeping a team aligned. The report found that 85% of respondents agree AI has shifted the bottleneck from writing code to reviewing and validating it. You've sped up input; you haven't changed throughput.\n\nThe accountability problem underneath this is structural. The report defines \"AI accountability\" as the ability to answer three questions about any line of AI-generated code: where it came from, what it was meant to do, and who owns it post-deployment. Most organizations can't answer these questions. Forty-three percent of respondents say they can't reliably distinguish AI-generated code from human-written code in their own repositories. Thirty-nine percent say their systems don't track code origins at all. The code is in production; no one knows how it got there or what would break if it needed to change.\n\nThis is where the slopware problem and the governance problem converge. The combination of faster generation, lower average quality, and reduced traceability produces a codebase that is simultaneously harder to review and harder to reason about — and the review queue is the thing everyone is already behind on.\n\n**Why it matters:**\n\n**For ICs:** Faster generation means your review queue is going to grow faster than your review capacity. Treat code review as a constraint to protect, not a formality to compress.\n\n**For leaders:** Measure delivery cycle time end to end, not just commit frequency. If your metrics only capture the AI-accelerated parts, you're reading a dashboard that flatters the wrong things.\n\n**For founders:** Before you scale AI-assisted development, ask whether your review, testing, and ownership infrastructure can absorb the output. If it can't, you're creating a maintenance debt at scale.\n\n## The Verdict: Real or Hype?\n\n**AI code verbosity and structural debt → Real.** The metrics are in: agent-generated code is measurably worse on the dimensions that matter for long-term maintainability.\n\n**AI as a full-stack delivery accelerator → Hype.** It speeds up the terminal; it does nothing for the organization surrounding it.\n\n**The deletion and governance gap → Real but early.** The problem is understood; the tooling and culture to address it are still catching up.", "url": "https://wpnews.pro/news/the-code-is-piling-up-nobody-budgeted-for-the-cleanup", "canonical_source": "https://fromtheterminal.substack.com/p/the-code-is-piling-up-nobody-budgeted", "published_at": "2026-09-15 00:58:00+00:00", "updated_at": "2026-09-20 19:53:25.665778+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "ai-research", "mlops"], "entities": ["Earendil", "Vicki Boykis"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/the-code-is-piling-up-nobody-budgeted-for-the-cleanup", "markdown": "https://wpnews.pro/news/the-code-is-piling-up-nobody-budgeted-for-the-cleanup.md", "text": "https://wpnews.pro/news/the-code-is-piling-up-nobody-budgeted-for-the-cleanup.txt", "jsonld": "https://wpnews.pro/news/the-code-is-piling-up-nobody-budgeted-for-the-cleanup.jsonld"}}