As AI agents take on more and more of the execution, the scarce resource shifts. It is no longer production capacity that is the bottleneck. It is the cognitive demands on the human: understanding enough of what is happening to steer, judge, and make the right decisions.
Agentic AI is often described as a productivity technology. That is true. For larger tasks, it is often not about giving a single agent full responsibility, but about letting several agents work on different parts of the process: writing code, running tests, debugging, reviewing results, and suggesting next steps. With increasingly powerful tools, tasks that recently required an entire software development team can sometimes be explored or taken far by a single developer with a laptop, a few prompts, and a set of agents.
But the story does not end there. As agents take over more of the execution, the human role changes. It does not disappear. What matters most is no longer primarily technical skill or typing speed, but the ability to hold the overall picture together: to understand enough of what is happening to steer the process, weigh alternatives, and decide what actually matters.
“We’re using old interfaces for new technology,” says Laurynas Adomaitis, AI researcher at RISE.
From Writing Code to Managing Production
The idea of the “dark software factory”, a factory so automated that no people need to be on the floor, has started to appear in AI circles as a metaphor for what software development may become. OpenAI has, for example, released Symphony, an open specification that turns an ordinary issue tracker into a kind of control plane for coding agents.
“You manage the production of software at a high level. You give it tasks. Then the agents write the code, review it, test it, and do the rest,” says Laurynas Adomaitis.
This is already happening today. Coding agents build prototypes, debug systems, and work directly in development environments. At RISE’s Mimer AI factory, startups regularly arrive after having “vibe-coded” a working product, and teenagers are building apps that make money in the same way.
But the factory metaphor can also mislead. A physical dark factory produces standardized things. Knowledge work is different. It is not only about execution, but about a continuous flow of choices: what should be built, how should the goal be interpreted, and which trade-offs are reasonable?
“Every non-trivial project runs into complexity,” says Sverker Janson, Director of the Center for Applied AI at RISE. “There are many choices and directions, and I need to hold the agent’s hand and guide it.”
The New Bottleneck
A single coding agent is manageable. Several agents running in parallel can be powerful. But somewhere along the way, the human becomes the bottleneck.
“I have seven parallel agent terminals open right now,” says Sverker Janson.
Laurynas Adomaitis recognizes the pattern.
“People do that. It works. But as an interface, it is very poor. It visualizes nothing. You do not know what the different agents are doing, where they are in the process, or how they relate to one another.”
Today’s interfaces, chat windows, command lines, and issue trackers, are inherited from older ways of working. They were not built to supervise several semi-autonomous systems that act, create, and coordinate at the same time. Agents can produce more information, more alternatives, and more half-finished results than a human can absorb. The scarce resource is no longer how fast the system can produce, but how quickly a human can understand, evaluate, and redirect what is being produced.
Interfaces Need to Keep Up
There is a concept in philosophy called affordances: interfaces make certain actions simple and natural, while others require more effort or become invisible. The interface you choose therefore shapes not only how you work, but also what you build.
Laurynas Adomaitis has explored how real-time strategy games can work as a metaphor: agents are represented spatially and can be grouped, stopped, and started with simple commands. The interface does not need to look like a game. The point is that if agentic AI is to become a tool for large-scale knowledge work, people need to move smoothly between levels, from high-level goals to concrete results, from strategic choices to direct intervention.
The next generation of AI interfaces may become less like a chat window and more like an environment for thinking, delegation, and control.
Trust Requires Deliberate Margins of Error
If you cannot review every action an agent takes, how can you trust the system?
One tempting answer is to require complete validation at every step. In practice, that can bring the entire process to a halt. Laurynas Adomaitis describes an experiment where several agents were allowed to run for two weeks with the goal of producing a substantial software product more or less from scratch. The result was only achieved once the agents were given room to be imperfect. When every step had to be fully validated, the system got stuck in loops and could not move forward.
“If the code is 95 percent correct, commit it, and it will fix the rest later,” says Laurynas Adomaitis.
This is not an argument for blind trust. It is an argument against the idea that meaningful control requires inspecting everything. We already work this way in teams: we delegate, make assumptions, discover errors, and correct them. What matters is not that every intermediate step is flawless, but that the process can be followed, adjusted, and steered back toward the goal.
What Kind of Loop Should the Human Be In?
The debate about agentic AI often focuses on what can be automated. But the more interesting question is what becomes more important for us humans when execution is no longer the scarce resource.
Perhaps it is about framing problems, setting direction, interpreting partial results, and protecting time for reflection. Well-designed systems can provide real leverage: making it easier to explore alternatives, build prototypes, and create tools adapted to a specific context. But that requires interfaces, organizations, and ways of working to be designed with that purpose in mind.
“We need positive visions that we genuinely try to make happen,” says Sverker Janson, “rather than being dragged, kicking and screaming, into whatever situation the economic forces create.”
The design question is urgent. Not: how do we remove the human from the loop? But: what kind of loop should the human be in?
Want to know more? Contact Laurynas Adomaitis or Sverker Janson at RISE.
