When every developer becomes a manager
I have been having versions of this conversation with researchers at conferences, and with colleagues in CS at UdS, for a while now. I thought I should finally summarize the point here, partly to keep the conversation around generative AI, agents, and education moving.
A lot of the discussion around generative AI in education has focused on whether we should allow it, whether it makes students lazy, and whether the work it produces is good enough.
I think we may be missing a different problem.
Generative AI is turning people into managers much earlier than before.
I do not have a complete answer to what this should mean for CS education. What follows is a collection of observations and questions that have been accumulating for a while, and that I think are worth putting down.
Delegation before experience
Working effectively with an AI agent is not just a matter of knowing how to prompt it. You have to explain the work to be done, provide the right context, break a larger objective into appropriate smaller tasks, anticipate where misunderstandings are likely, decide what needs to be checked, and verify that the result actually meets the original objective.
These are remarkably similar to skills we normally acquire through managing work.
Historically, most of us learned them gradually. We first did the work ourselves. We watched supervisors, research advisors, project managers, and team leads. We experienced good delegation, bad delegation, and sometimes no delegation at all. Eventually, some of us chose to take responsibility for other people's work and learned, often through mistakes, how much ambiguity an instruction could create, what information someone would need to succeed, and what needed to be checked along the way.
Becoming a manager was, to a large extent, a professional decision. If you liked that kind of responsibility, you could move toward it. If you did not, you could remain an individual contributor and become excellent at your craft without necessarily having to manage the work of others.
Agents are significantly changing that distinction.
Anyone, including BSc students or even high-schoolers, can now delegate substantial work to several agents without ever having managed anyone, supervised a project, or developed much intuition for what can go wrong when work is poorly specified or insufficiently verified. More importantly, they may not have chosen to become managers at all. They have not necessarily been trained for that role, nor are they necessarily thinking of themselves as responsible for managing delegated work well. Management-like responsibilities are simply becoming part of doing the work.
What made me think more seriously about this was hearing managers complain about how uncritically some new hires, fresh from university, delegate work to AI. The problem is often not simply that they use AI, or even that the AI produces something wrong. It is that they do not seem to anticipate what could go wrong with the task they have delegated or how they should verify that the work was actually done to the expected standard.
I have noticed the same pattern with students, at different levels.
The appeal of agents is obvious: they can do work for you. The most appealing interaction would be to hand over a task with minimal explanation and simply trust the result. But that is not where we are in 2026. Effective delegation still requires careful specification, context, supervision, and verification.
In my experience, people with management experience often approach agents differently. They tend to treat delegation as something that needs to be actively managed, rather than as a one-shot instruction.
What this means for CS education
That makes me wonder whether we are framing the educational challenge of AI in computer science too narrowly. So far, much of the discussion has focused on how generative AI should be handled in teaching and assessment: whether students should use it, under what conditions, and how lectures, assignments, and exams should adapt. At the same time, curricula are evolving to teach more AI itself, through new courses on topics such as agentic AI and by integrating generative AI tools into more classical CS subjects such as software engineering.
Both are important. But there may be a third part of the problem that receives less attention: if students are increasingly going to delegate meaningful work to agents, we also need to think about whether we are teaching them how to delegate, supervise, and evaluate that work effectively.
Students certainly need to understand large language models, AI tools, and how generative AI can be used in software development. But perhaps we also need to teach delegation, task decomposition, specification, supervision, and verification much earlier.
Not because every computer science student should become a people manager. But because, increasingly, even an individual contributor is managing work.
The traditional progression was roughly this: learn to do the work, become proficient at it, and then, if you chose that path, gradually take responsibility for other people's work and learn how to manage along the way.
Agents are collapsing part of that progression. And this probably is not unique to computer science. Any field in which agents begin doing meaningful chunks of work may face the same problem.
This is the shift I think CS education needs to take seriously. We may be raising the first generation of professionals for whom managing work is no longer a career choice, but part of being an individual contributor.