80% AI-Written Code. So Why Is Your Codebase Getting Dumber?

OpenAI's president went on record this week saying AI now writes 80% of your code. Business Insider ran it as a headline. Half of tech Twitter took a victory lap. Nobody asked the only question that matters to a CEO: who is checking the 80%, and what happens to your company when the people who used to check it forget how.

I run a tech company in Jönköping building software with AI agents every day. We use Claude Code. We use agent harnesses. We ship fast. So this isn't a Luddite take. This is a builder telling you the honest version of what's happening inside the machine, because right now the machine is quietly moving engineering judgment out of your senior team's heads and into a black box, and almost nobody is pricing that risk correctly.

The 80/20 Flip Nobody Is Pricing Correctly

Here's the flip in plain terms. Five years ago, engineers wrote 80% of the code and reviewed AI suggestions for the other 20%, mostly autocomplete-level stuff. Today that ratio has inverted. AI generates the bulk of new code across serious engineering orgs, and humans are supposed to be reviewing it.

The problem is that review is not happening at the rigor the old writing process forced on people. When you write code yourself, you build a mental model of the system as a side effect. You can't write a race condition fix without understanding the race condition. When you review AI-generated code, especially under deadline pressure, you're pattern matching for "does this look right," not building understanding. That's a fundamentally different cognitive task, and it's worse at catching subtle failure.

Anthropic's own research, out this week, flags exactly this: AI assistance is measurably eroding how junior engineers build coding skill. Not a rival's talking point. Anthropic saying it about their own product's second-order effect. That should make every CTO reading this sit up. If the company selling you the tool is warning you about the skill decay it causes, believe them.

Meanwhile, SD Times is reporting something the hype cycle conveniently skips: companies are actively pulling back on AI coding tools. Not because the tools don't work. Because the tools work exactly as advertised, and what got advertised was velocity, not judgment. Velocity without judgment is how you ship a fast car with no brakes.

Vibe-Kodning Is a Swedish Problem Too

Swedish tech media has fallen into the same trap as everyone else. Unite.AI's "Bästa AI-kodgenererare" listicle is getting shared in every founder group chat in Stockholm right now, and it reads like a menu, not a warning label. Meanwhile Towards Data Science published a genuinely important piece on hidden technical debt in AI-generated systems, and it's getting buried under the "vibe-kodning" excitement. That's backwards, and it's a very Swedish kind of backwards: we love adopting a trend efficiently, and we're slower to ask what it costs later.

Here's what I see from Jönköping versus what I hear out of San Francisco. In the US, the dominant posture is "ship the demo, raise the round, worry about debt after Series B." That's a legitimate strategy if you're optimizing for a 18-month fundraising story. It's a terrible strategy if you're optimizing for a company that still exists in five years.

In the Nordics, we don't have the same appetite for that gamble, mostly because our capital markets don't reward pure velocity theater the way American VC does. That should be our advantage. Instead, half the founders I talk to at Sweden's startup events are running the exact same vibe-coding playbook, just with a smaller marketing budget and less runway to survive the technical debt bill when it lands.

DI has covered how Swedish scaleups are under real pressure to show AI adoption to investors, and that pressure creates exactly the wrong incentive: adopt AI coding tools fast, show the velocity number, skip the boring infrastructure that makes velocity sustainable. I get why founders do it. I think it's the wrong bet.

The Boring Pipeline Wins in 18 Months

Here's my actual prediction, and I'll put a date on it. The Swedish and Nordic companies quietly building review pipelines, test gates, and architectural guardrails around their AI coding agents will out-ship the vibe-coding shops within 18 months. Not because they'll write more code. Because they won't have to spend Q2 2027 unwinding a codebase nobody understands.

What does a boring pipeline actually look like? It's not exciting, which is exactly the point.

1. Mandatory human architecture review before agent execution

The agent doesn't decide the system design. A senior engineer decides the shape, the boundaries, the data model. The agent fills in implementation inside constraints a human already understood. This is the difference between AI as junior engineer and AI as architect. Never let it be the architect.

2. Diff review with actual understanding, not approval theater

If your review process is "click approve on the PR," you don't have a review process. You have a rubber stamp with a human name attached. Real review means the reviewer can explain, out loud, why the code works, not just that tests pass.

3. Test coverage that catches intent, not just syntax

AI-generated code passes AI-generated tests at a suspiciously high rate. That's not confidence, that's the model checking its own homework. You need tests written or reviewed by someone who understands the business logic, not just the code path.

4. A skills budget for your juniors

This is the part nobody wants to hear. If your junior engineers only ever review AI output and never write from scratch, you are burning your future senior bench. Anthropic flagged this as real, measurable erosion. Build deliberate practice into your process even when it's slower. It's an investment, not overhead.

This is exactly how we approach AI agent development at HEIMLANDR. The agent is fast. The pipeline around the agent is what makes the output trustworthy. Speed without a pipeline is just a faster way to accumulate debt you can't see yet.

Where This Goes: AGI, Regulation, and the Judgment Gap

Push this trend five years out and the stakes get bigger, not smaller. As models get closer to general capability, the temptation will be to remove humans from the loop entirely, not just from writing code but from reviewing it. Full autonomous agent pipelines, code to deploy, zero human eyes. Some companies are already experimenting with this. It will work beautifully right up until it doesn't, and when it fails it will fail in production, at scale, with nobody who understands the system well enough to diagnose it fast.

The EU's AI Act touches high-risk AI systems, but it has almost nothing to say about AI-generated software as a systemic risk category. There's no regulatory framework anywhere, not in Brussels, not in Washington, that treats "AI wrote the code your bank runs on and nobody reviewed it properly" as the risk category it actually is. That's a genuine policy gap. Sweden and the EU move carefully on AI regulation, which is usually a strength, but carefully should not mean blind to a risk that's compounding in production systems right now.

My honest prediction: within 3 to 5 years we'll see the first major infrastructure failure, financial or otherwise, traced directly back to AI-generated code that nobody with real understanding ever reviewed. When that happens, expect a regulatory overcorrection, probably mandatory audit trails for AI-assisted commits in critical infrastructure. Companies that already have boring review pipelines will comply in a week. Companies running pure vibe-coding will need six months and a very uncomfortable conversation with regulators.

The path toward AGI doesn't remove this problem. It sharpens it. More capable models mean more confident-looking output, which means human reviewers get lazier faster, which means the judgment gap widens even as the code quality superficially improves. The skill that matters most in an AGI-adjacent world isn't prompting. It's the ability to look at a system and know, structurally, whether it's sound. That skill takes years to build and minutes to atrophy.

What to Look At

If you're serious about building the boring pipeline instead of the vibe-coding shortcut, here's where I'd start looking this week:

• ECC (Skills, instincts, memory), an agent harness performance system built specifically around research-first development for Claude Code and similar tools. It's the kind of infrastructure that treats the agent as a component in a system, not the whole system.
• opencode, an open source coding agent worth studying if you want to understand what's happening under the hood instead of trusting a closed vendor black box.
• n8n for anyone building agent workflows with actual human checkpoints wired into automation, rather than fully autonomous pipelines with no gate.
• Anthropic's own published research on skill erosion, worth reading directly rather than through a headline. It's the most honest document a vendor has published about their own product's downside this year.

What You Should Actually Do This Week

Stop measuring AI adoption by lines of code generated. Start measuring it by defect rate reaching production, by mean time to understand an incident, by how many of your engineers can explain the system they're shipping without opening a chat log. Those numbers tell you if you're building a company or building a liability with a nice demo.

If you're a founder trying to move fast without gambling the codebase, this is exactly the gap between Rapid MVP done properly and vibe-coding dressed up as MVP work. Speed to market and engineering judgment are not opposites. They only look like opposites when your pipeline is missing the review layer. We build fullstack systems with that layer baked in from day one, because retrofitting judgment into a codebase after the fact costs ten times what building it in costs upfront.

Sweden has a real shot at leading here, not by rejecting AI coding tools, but by being the market that takes the boring parts seriously while everyone else races for the demo. That's not a small ambition. That's the actual competitive edge available right now, and almost nobody is taking it.

Fredrik Brunnberg is the CEO of HEIMLANDR.IO, building AI and software solutions from Jönköping, Sweden. This is the daily HEIMLANDR briefing. If you found this valuable, share it with someone who builds things.