Understand

Watching your AI agents work

The main barrier to delegation is not distrust of the machine: it is invisibility. You hand a task to a system that shows nothing, and you find yourself refreshing a board to guess whether it is still working. Most leaders give up there — not because the result was poor, but because they never managed to picture what was happening.

1. Why seeing changes delegation

Running a human team relies largely on information nobody puts into words. You see who is present, who talks to whom, who looks overwhelmed, who has done nothing for two days. None of these signals appears in a report, and yet they are what triggers intervention at the right moment.

Agent systems remove all of that at once. What remains is a task queue and technical logs: enough to check afterwards, useless for steering during. This explains a common phenomenon — companies equipped with agents that they barely watch, having no idea what they are doing.

Making the work visible is therefore not a matter of pleasantness. It is what allows you to delegate more, because you stop fearing what you cannot see.

2. What the office shows

In nullbot, the company appears as an office where each agent occupies a desk. At a glance you read what no table delivers as fast.

Who is working, and on what — each agent at their desk, with the current task displayed above them.
Who is idle — inactivity is immediately visible, whereas it is almost invisible in a task list.
Who talks to whom — exchanges between agents, delegations, the approval requests waiting for you.
Who reports to whom — the chain of command, as it has been defined: who can hand work to whom, and who is accountable.
Meetings — several agents gathered around a subject, with the discussion readable and the decisions that come out of it.

None of this is decorative simulation: these are the real states of the system, laid out in space rather than in a table. An agent who does not appear at their desk is genuinely not working.

3. Looking inside an agent

The office shows activity. A second view shows competence: what the agent has learned, what it can do, what it retains from one task to the next.

It is a more serious question than it appears. An agent that starts from zero every morning costs a lot and progresses little; an agent that accumulates memory on your files becomes more useful month after month. Being able to observe that progression — rather than assume it — changes what you are prepared to entrust to it.

It is also what makes the system explainable to a team. "The agent knows our qualification rules because it learned them here, and this is what it retained" can be discussed. A black box cannot be discussed: it is endured or refused.

4. Seeing is not proof

The limit has to be stated plainly, because that is exactly where spectacular demonstrations mislead.

A visual representation serves to understand and to intervene. It does not serve as proof. What counts as evidence remains the audit log: which agent did what, at what moment, on what data, at what cost, and who had authorised it. That register is what you produce in the event of a dispute, an inspection or a simple internal disagreement — not a screenshot of an office.

Both are necessary and neither replaces the other. A system that shows nothing cannot be steered; a system that shows without recording cannot be defended.

5. Frequently asked questions

What is a 3D office for in business software?

To answer in one second a question that otherwise takes several clicks: what is my company doing right now? The spatial view is not decoration — it exploits the fact that a human spots an anomaly in a scene far faster than in a list. A conventional two-dimensional view remains available for those who prefer it.

Does it slow the machine down?

The view is automatically lightened on modest configurations and on mobile, and it pauses when off-screen. The agents' work does not depend on the display: closing the window stops nothing.

Can I step in from the office?

Yes. You talk to an agent, hand it a task, take back control of what it is doing, approve or refuse what it proposes. That is the point: seeing without being able to act would only provide a feeling of control.

How many agents can work in it?

Five on the Solo plan, unlimited beyond. The useful number depends above all on the work genuinely delegated: ten agents without a clear task produce less than two well-framed ones.