The Risk of Agency: How AI Forces Us to Take It, and Why Germany Will Suffer

Julian Habekost

Jul 08, 2026

As If Superintelligent Aliens Landed on Earth and Willfully Enslaved Themselves to Us

It is really, really hard to communicate to outsiders what kind of revolution just happened in the world of software engineering. Sure, people make hyper-personalized birthday greeting cards with AI, they let the AI word their emails, their letters, and let AI help them navigate bureaucracy. But nothing comes close to typing in a four-sentence instruction and then watching the AI agent for ten minutes reading the source code, searching the web, arguing with itself about the code, starting to change the code, programming tests to hunt down bugs, finding the bugs, correcting itself, and lastly documenting everything in a manner that I really should have done, but in reality, none of us ever did. What the AI did there in ten minutes would have been a day’s work a few months ago. And that concerns technical and coding domains that I am an expert in. It feels as if superintelligent and super-knowledgeable aliens landed on Earth and have willfully enslaved themselves to us. But without all the ethical implications.

Please Don’t Ask Me to Review Your AI’s Code

As a senior engineer, it is part of my job description to review the code of juniors. But with AI, I am increasingly reluctant to do that, because their code is now also, just like mine, written by AI. I already review my AI’s code; if I also review your AI’s code, what is actually your job then? The relationship between an engineer and an AI coding agent is very much like the relationship between a senior and a junior (albeit the best junior that ever existed). There is no point anymore in tying up junior-sized packages of well-specified tasks to give to juniors so they can feed them to an AI, when instead I can simply do it myself with less communication overhead.

No wonder the job market for junior software engineers has been crushed recently. If everyone has access to AI juniors now, every one of us human engineers has to become a senior. We can discuss the technical direction, but I am not taking responsibility for your AI’s code; you have to do that yourself, or you’re obsolete. If your AI’s code breaks something, I’m going to shame you (appropriately for a work relationship) in the next meeting for it.

To be fair, I work in corporate research and in experimental innovation projects with a lot of prototypical, alternative variants; we move fast and break things a lot anyway. If it happens, it usually means one to three engineers have one to three bad days at work. It’s a risk we can afford to let everyone on the team take.

This makes us one of the biggest beneficiaries of agentic AI engineering, because we can afford to go full YOLO, as the kids would say these days. Instead of exploring two different variants with two seniors and five juniors, we can explore up to seven variants now at the same time. I am expecting direction and ideas from everyone, and this is exactly what it means to act out agency.

The Crisis of the Junior is Not About Skills, It is About Denied Grassroots Agency

The closer you get to production code, to established businesses and proven software that needs incremental updates rolled out to users frequently, the less reasonable it becomes to develop seven different alternatives. Even there, there is an opportunity now to try to redevelop modules and parts of the software with much lower costs, but that again would just be innovation that is not rolled out immediately. For maintaining production code, someone has to make the final calls, and this someone will be the same person as before; but now that person also has AI juniors at their disposal. There is hardly any use for actual human juniors acting like seniors in production environments, hence almost no use for human juniors at all anymore in software maintenance.

A lot of people think this is a skill progression issue; they think that it is impossible to become a senior without having been a junior first. But firstly, I am not talking about skill; I am talking about roles and the agency they are granted. And secondly, I also disagree.

AI agents actually reduce the learning curve of coding and technology drastically because you can literally ask them to explain code or technology. We had people join our project who had no Deep Learning experience that went on to contribute an actual competitive neural network from scratch and also were able to defend its design decisions in a theoretical discussion—because they had done their homework and discussed it with their AI agent, and not just piped its code through. My brother, a teacher of history and music, has started to learn coding after discovering agentic AI development. He successfully asked the AI to develop a bookkeeping software for his DJing side hustle and also realized that with even a little bit of programming knowledge, the impact on the agent and its result is huge. Before that, learning to code was a long journey before you could achieve anything useful; now, the tiniest seed of skill can be leveraged into raising your ceiling. The learning curve harvest is now better than it ever was.

There is no problem with skill progression; there is a crisis of growing grassroots agency, and the fact that companies are not designed to allow so many people to have so much of it, because agency undeniably comes with risks.

The More Possibilities Humans Have, The Bigger Their Agency and The Bigger Its Risk

I define the risk of agency as the risk, specifically the opportunity cost risk, of human action (and even deliberate non-action) in the world, as an individual or as a group. Do you want to repair something or invent a substitute from scratch? Do you want to go to medical school, law school, or rather pick up a trade? Do you buy a bigger house or go on more vacations? Do you spend more time with your kids or earn more money as a tradesman so they can go to medical school? Does a company research X or Y, or rather save the money and piggyback on others doing the research? The risk of agency is a risk that by definition can conceptually never be overcome; for every risk you mitigate and every insurance invented, new doors open up that carry new risks of agency. The irony is that the more possibilities humans have, the bigger the agency and the bigger its risk.

Let that sink in; and let me give you a crass example: Twelve-year-old Anne Frank in 1941 had almost no risk of agency, while my risk of agency as a similar-aged coding wunderkind in peaceful 2003 Germany was moderately high. All agency was stripped away from Anne Frank by the Nazis, and without agency, there is no risk of agency. (And while I performed maybe satisfactorily given my starting position, Anne Frank had the biggest possible imprint on the world given the little agency remaining that the Nazis could not take from her.) But there was nothing she could do about her fate.

So the quintessence is: only people with opportunities have opportunity cost risks. AI, specifically agentic AI software engineering, changes the risk structure from two sides: it gives us all a lot more opportunities and a lot more room for agency, but it also makes a lot of things much easier and faster; hence, the risk-reward curve of agency gets much steeper. Things that last year would have been classified as low-to-moderate risk are now fast and sometimes even trivial to try. If everyone can do it, it is not that economically rewarding, at least not in isolation. Finding and building the things that matter is now a much more holistic task of understanding both the problem and the technology to a deeper level than everyone else with access to an AI, and of verifying it with the actual audience. It is almost a bit more like writing a script for a movie now.

AI gives humans agency; it net-positively creates and democratizes opportunities, but that translates to a steeper risk-reward profile of agency. Corporations can only profit if they can push their risk tolerance, if they can redirect the risk to those employees with agency, and if they can find hybrid solutions with a little more of a startup’s risk-reward feedback.

Execution and Half of the “How?” Are Now Solved

The classic hierarchy in large corporations is: execution at the bottom, “how?” at the middle, and “what?” at the top. The problem is, for software engineering, execution and half of the “how?” are now solved by AI agents; what is left is the half of the “how?” that is harder and more connected to outside factors and stakeholders the AI does not understand, along with the “what?”, the core question of human agency.

The “what?” in software engineering is the CEO’s job in a tech startup; it is connected to everything that matters for the mission: the customer’s problems or desires, the market and competition, and technical feasibility—the still-unsolved part of the how. People often think that the CEO of a tech startup is only responsible for customers, the market, and sales, while the CTO is driving all the technical decisions. That might work for something like Airbnb (I would hardly call that a tech startup), but not for an actual (deep) tech startup like Anthropic. When you are pushing deep tech innovation, technical feasibility and market perspectives are two sides of the same coin; trying to separate the tech and sales responsibility in innovation onto two shoulders is like putting two drivers in a cockpit of a race car and giving one the gas and the other the brake pedal. It is not going to work; it does not matter which of them gets to hold the steering wheel.

The role of the deep tech startup CTO is thus highly misunderstood; it was never meant to be more than a chief high-tech janitor. To be fair, there are many CEO/CTO pairs who actually act like tandem CEOs, and that is a completely different story. Even with a slightly different focus, both understand all mission-critical tech and market factors in a way that they can both ride the tandem (say its steering is linked for this metaphor’s sake) alone, and they are both responsible for the whole tandem.

We All Have to Become CEOs of Our Projects

With agentic AI juniors doing the execution for us, both the role of the engineer and the role of the manager will converge towards the same CEO-ish role, a project lead role. They are converging from different sides, and the issues and opportunities that will arise for them are different. A classic project manager who is not able to instruct and discuss with an AI agent to further develop the prototypical deep tech software they are heading clearly did not have the technical depth necessary for the job anyway; they were probably carried by senior engineers doing big chunks of their job before. But on the other hand, engineers will have to learn to validate the impact of their ideas, and to discuss with potential customers and stakeholders. They have to overcome the typical engineer’s diseases of reinventing the wheel, ‘not-invented-here’ syndrome, ‘people can’t comprehend my genius’, and ‘it depends’—and not just by making a pinky promise, but by bearing the sheer weight of the new responsibility. They have to assume the risk of their agency.

This is what I love about the AI situation: it exposes laziness and lies (including those told to themselves) of both sides of the engineer versus management clash, because we will soon all be alike. What was once execution at the bottom, “how?” at the middle, and “what?” at the top will become “what?” at every level, but with different-sized budgets and risk profiles. Or rather should become, because the obvious question is: can the hierarchy even project or align the risk onto the bearers of the specific responsibility?

The Risk of Agency Is Shattered in Big Corporate Hierarchies, Leading to Inadequate Actions

When people ask me what I do in my job at Bosch, I usually jokingly say: Hop

[truncated for AI cost control]