Navigating what we cannot fully see

The article In the control room for AI agents grew out of a conversation about how agentic AI is changing the nature of human work. This blog post continues that conversation, but turns toward a more fundamental question: what does it mean to be in control when work unfolds inside systems no human can fully oversee in detail? 

Being in control 

When we cannot see what is actually happening, it is easy to feel that control has slipped away. We cannot follow every decision, inspect every step, or see how the process is developing. The result can be a sense of helplessness in the face of something that keeps moving without us. 

A useful concept here is absorbed coping, coined by Hubert Dreyfus in early 1990s. It describes our ability to act skilfully in a situation without first analysing every detail. 

We do this all the time. An experienced project manager cannot know everything happening inside a complex project, but can still sense when things are moving in the right direction — or when something starts to feel off. A leader cannot read every email in an organisation, but can still make decisions based on patterns, signals and experience. 

In other words, we can trust a process without giving up control. Not blind trust, but a practical feel for when things are working, when they need adjustment, and when intervention is needed. 

That is not the absence of control. It is another form of control: a fast, competent way of handling situations we cannot fully oversee. 

But absorbed coping should not be the whole ambition. Managing a process is not the same as understanding every part of it. If we are designing the environments around AI agents, the goal should not be to make humans absorb even more complexity. The goal should be to build control rooms that make the right things visible, the right decisions possible, and the right pauses available. 

Seven terminals and no map 

In our conversation, Sverker Janson noted that he can have seven parallel agent terminals open at the same time. That is no longer unusual. Many people working with agentic AI systems let several agents run in parallel, check in occasionally, adjust when something looks wrong, and intervene when the process starts to drift. 

It works. But as an interface, it is remarkably poor. 

You cannot really see what the different agents are doing. You do not always know where they are in the process. You cannot see how they relate to one another, or how a decision in one terminal affects work in another. You are navigating a complex, dynamic system with tools designed for a sequential workflow: one person, one task, one terminal, one line at a time. 

So the question is not whether a human can see everything. The question is what the human needs to see, when they need to see it, and at what level of abstraction. Control cannot mean total access to every detail of the complexity AI systems are handling. It means having signals good enough to act on. 

Allowing errors to enable more autonomous processes 

We also discussed an experiment in which a multi-agent system was allowed to work for two weeks with minimal human involvement. The goal was to produce a larger software product. The interesting part was not only what the system managed to do, but what caused it to get stuck. 

When every intermediate result had to be perfectly validated before the next step could begin, the system quickly fell into loops. It became cautious, repetitive and slow. It only started moving once the agents were allowed to be imperfect: to produce code that was good enough to continue, and then fix shortcomings along the way. 

That may sound risky. But it is also quite human. 

Control cannot be reduced to manually checking every micro-step. It has to be designed into the system through the right feedback, the right stopping points, the right levels of abstraction, and the right forms of responsibility. 

So the question is not whether we should accept errors. The question is how we make sure they are discovered and corrected along the way. 

New interfaces with room for reflection 

An interface does not only shape how we work. It shapes what we see. And what we see affects what we are able to understand, question and imagine. 

In the conversation, we also returned to the idea of affordance in decision-making. Some decisions are easy to make. They require little context, little reflection, and can be handled quickly. Others ask more of us. They need time, context and room for reflection. One could say they have a higher affordance. 

This matters deeply for the design of AI agent interfaces. Humans should not be forced to stop at every micro-step. But they must be able to stop when a decision actually requires it. A good control room should help us distinguish between what can continue to flow, what needs adjustment, and what calls for real thought. 

Terminals are fast, precise and powerful. But they usually show one process at a time, and they lack many of the cues humans use to understand complexity: spatial relationships, overview, status, direction, dependencies and change over time. 

That is why we discussed real-time strategy games as a metaphor. Not because agent interfaces should look like games, but because these games have solved a real design problem: how do you give a human an overview of a complex system with many moving parts, while still allowing them to act at the right level, at the right moment? 

You do not need to control every unit all the time. You need to know where attention is needed. You need to be able to zoom out, zoom in, prioritise, delegate, stop and redirect. 

That is what the next generation of agent interfaces will need to support: not just faster execution, but better movement between overview, action and reflection. 

The human loop needs to be redesigned 

When the process is running, the agents are doing what they should, and nothing seems to need attention, the human should not have to micromanage. Their attention should be available for something more valuable: interpreting context, judging significance, weighing risks, noticing when something is wrong, and taking responsibility for direction. 

But when something unexpected happens, the human must be able to move into a more reflective mode. They need to understand what happened, what options are available, and what consequences those options may have. The human loop is therefore not about approving every micro-step. It is about being able to enter at the right level, at the right time. 

The question is not how we remove the human from the loop. The question is what kind of loop the human should be in. 

If agents only increase the pace, they may also increase fragmentation: more output, more decisions, more oversight. But if these systems are well designed, they can instead give humans better leverage — more room to explore alternatives, frame problems and steer direction. 

AI agents have already taken over a considerable part of execution. But that does not make humans less important. It makes the design of their control rooms more important. 

Continue the conversation 

Laurynas Adomaitis is an AI researcher at RISE, focusing on agentic systems and human–AI interaction. Sverker Janson is Director of the RISE Centre for Applied AI. 

This text builds on the reasoning in the article In the control room for AI agents, where we explore how AI agents are changing the human role in work. For more reflections on applied AI, research and societal change, follow the AI at RISE newsletter.