Your Developers Are Now Managers. Nobody Trained Them.

Every Developer Just Got Promoted. Most Will Fail.

Right now, today, Computer Sweden is running a piece titled "AI-kodning kräver att utvecklare blir bättre chefer". AI coding requires developers to become better managers. Read that again. The biggest Swedish tech publication is naming what Silicon Valley still dances around: your developers are no longer developers. They are engineering managers of a team of non-deterministic junior devs who hallucinate, skip security reviews, and have never read your internal docs.

And neither your org charts nor your toolchains are built for this.

I run HEIMLANDR.IO, a tech company in Jönköping, Sweden. We build AI agents, SaaS products, and fullstack systems. We are not observers. We use AI coding tools daily. We ship production code with them. And I am telling you: the problem is not that AI writes code. The problem is that nobody taught your people how to manage something that writes code confidently, incorrectly, and at scale.

The Management Gap Nobody Talks About

Think about what happened to your senior engineer in the last twelve months. Before, she wrote code. She reviewed PRs from humans who understood context, had read the internal docs, and feared the wrath of the security team. Her job was craftsmanship.

Now her job looks like this: prompt an AI agent. Read the output. Decide if it is right. Check if it introduced a vulnerability. Verify it did not hallucinate an API endpoint that does not exist. Confirm it respected the architecture decisions documented in a Confluence page from 2024 that the AI has never seen. Repeat forty times a day.

That is not development. That is management. Specifically, it is managing an employee who is fast, eager, confident, and frequently wrong. An employee who never pushes back, never asks clarifying questions on its own, and never says "I don't know."

Security Boulevard is flagging this right now: your AI coding assistant has never read your security wiki, and it is already writing half your code. Half. Think about what that means for your attack surface.

The Toolchain War Is a Distraction

Microsoft and Google are racing to own the AI coding layer. CNBC covers it like a horse race. Claude Code vs Codex CLI benchmarks fill tech blogs. The Spotify veterans behind Lovable's competitor make headlines in Dagens Industri. Everyone argues about which model writes better code.

I do not care which model writes better code. Neither should you.

The model quality gap is shrinking every quarter. What matters is the supervision and review infrastructure around these tools. And almost nobody is building it. The real battleground is not generation. It is verification. It is governance. It is the organizational design that sits between "AI wrote this" and "this is in production."

At HEIMLANDR, when we do rapid MVP development, we use AI agents heavily. But we have built our own review loops. We have internal standards documents that we feed into context windows. We have humans who understand their job is not to write code anymore. Their job is to guarantee the AI's code meets our bar. That is a fundamentally different skill.

What the Actual Workflow Looks Like

Let me be concrete. Here is what AI agent development looks like in a well-run team right now:

1. Specification. A human writes a clear, constrained spec. Not a vague prompt. A spec. With acceptance criteria, edge cases, and explicit constraints.
2. Generation. An AI agent writes code against that spec. Fast. Cheap. Confident.
3. Review. A human reviews the output like a senior engineer reviews a junior's PR. Line by line. With suspicion. Checking for hallucinated dependencies, missed error handling, security holes, and architectural drift.
4. Context injection. The human feeds back corrections, internal standards, and domain knowledge the AI did not have. Another round.
5. Verification. Automated tests, security scans, and integration tests. Not optional. Mandatory before merge.

Steps 1, 3, 4, and 5 are management tasks. Only step 2 is "AI doing the coding." But most companies treat the whole thing as step 2. Prompt, accept, ship. That is how you get production incidents that no one can explain because the code was written by a system that cannot be asked "why did you do this?"

Sweden vs. The World: We Name Things Earlier, But Move Slower

There is something very Swedish happening here. Computer Sweden publishes a clear-eyed headline about developers needing to become managers. DI covers the infrastructure race with nuance. Swedish tech media is naming this problem faster than most US outlets, which are still stuck on "AI will replace developers" vs "AI will never replace developers" debates.

Good. But naming the problem is not the same as solving it.

Here in Jönköping, I see mid-sized software companies that still organize teams around code output. Performance reviews still measure lines of code, features shipped, velocity points. Nobody measures review quality. Nobody measures "bugs prevented in AI-generated code." Nobody tracks how well a developer manages their AI tools.

The US has a different problem. They are moving fast and hiring "prompt engineers" as if that is a real discipline separate from software engineering. It is not. A prompt engineer who cannot read code is useless. A software engineer who cannot manage AI output is becoming useless too. The answer is the same person, with new skills.

Asia, specifically China and South Korea, is moving fastest on AI agent development integration. But they are also producing the most catastrophic AI-generated code failures, because speed without supervision scales mistakes.

EU regulation, predictably, is somewhere between "concerned" and "drafting a framework." The AI Act exists but says almost nothing about AI-generated code quality, liability for AI-introduced vulnerabilities, or who is responsible when an AI agent writes code that causes a data breach. Swedish regulators at IMY are focused on GDPR implications of AI but not on software supply chain integrity. This is a gap. It will be exploited before it is closed.

Where This Goes: 2027-2030

Let me lay out what I think happens next.

2026-2027: The "manager layer" becomes explicit. Companies that survive will create formal roles and processes for AI code supervision. Job titles will shift. "Senior Developer" starts meaning "someone who can manage AI coding agents effectively." Companies that do not make this shift will ship increasingly unreliable software and will not understand why.

2027-2028: Liability frameworks emerge. The first major lawsuits over AI-generated code vulnerabilities hit. EU regulators scramble. Insurance companies start requiring AI code audit trails. Suddenly, every company needs to prove that a human reviewed AI-generated code before it hit production. Companies without those logs are exposed.

2028-2030: As we approach more capable AI systems, the management problem inverts. AI agents start managing other AI agents, with humans as executive oversight. The developer's role becomes something closer to a CTO of a small autonomous team. You are not reviewing individual lines. You are setting objectives, defining constraints, and auditing outcomes. This is where AI solutions architecture becomes the core competency, not AI usage.

The path toward AGI does not eliminate the management problem. It escalates it. The better these systems get, the more dangerous unsupervised operation becomes, because the mistakes become harder to catch and the blast radius grows.

What This Means for Your Org Chart

If you are a CTO reading this, here is the blunt version: your org chart is wrong. You have it structured as if humans write code and other humans review it. The reality is that AI writes code and humans need to supervise it with a completely different mindset, skill set, and toolset. Your 1:1s with senior engineers should include questions like: "How are you validating AI output? What did you catch this week? Where did the AI surprise you?" If those conversations are not happening, you are flying blind.

What to Look At

A few things worth your attention right now:

ECC (Engineering Copilot Controller) is trending on GitHub with over 200K stars. It is an agent harness performance optimization system built for Claude Code, Codex, Opencode, Cursor, and others. Skills, instincts, memory, security, and research-first development. This is exactly the kind of supervision layer I am talking about. Worth studying even if you do not adopt it directly.

Opencode is another one to watch. An open source coding agent sitting at 170K+ stars. The interesting thing is not the agent itself. It is the community building review and governance patterns around it. Open source supervision tooling is where the real innovation will happen, because no single vendor has the incentive to build robust oversight of their own AI.

n8n keeps growing as a workflow automation platform with native AI capabilities. If you are building internal review pipelines for AI-generated code, combining n8n with custom security checks and automated testing is a practical starting point. We use workflow automation like this internally at HEIMLANDR when orchestrating multi-step AI agent tasks.

What You Should Actually Do Monday Morning

Stop thinking about AI coding tools as productivity boosters. Start thinking about them as new team members who need management.

Concretely:

• Audit your AI code ratio. How much of your production code is AI-generated? If you do not know, that is your first problem.
• Formalize review standards for AI output. Different from human PR review. AI makes different kinds of mistakes. Train for those patterns.
• Update your performance criteria. Stop measuring developer output by volume. Start measuring review quality, defect catch rate, and AI supervision effectiveness.
• Feed your internal docs into context. Architecture decisions, security policies, coding standards. If it is not in the AI's context window, it does not exist for the AI.
• Create audit trails. Log what the AI generated, what a human changed, and why. You will need this for compliance. Sooner than you think.

The Uncomfortable Truth

Here is what I keep coming back to, sitting in Jönköping, watching the global software development scene from Sweden. We are in the middle of the biggest role transformation in software engineering history, and most companies are treating it as a tools upgrade. "We adopted Copilot. Productivity is up 40%." Great. Is quality up? Is security up? Do your developers know their job changed?

Software development in Sweden, and everywhere else, is no longer primarily about writing code. It is about managing systems that write code. The companies that internalize this first will build better, faster, and more reliably. The ones that do not will ship fast and break things in ways they cannot diagnose, cannot explain to regulators, and cannot fix without understanding what the AI did and why.

Your developers got promoted. Train them for the job they actually have now. Or someone else will hire them once they figure it out on their own.

That is the dispatch for today. Happy Nationaldagen. Go build something.

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.