The context switch should be the machine's problem

The Slack ping arrives while you are deep in something hard. It is a small question, two minutes of work, no big deal. You answer it. The two minutes are not the cost. The cost is the next twenty, the ones you spend climbing back into the problem you were holding, reconstructing the half-built model in your head that the ping knocked over. The interruption was tiny. The context switch it triggered was not.

This is the tax nobody puts on the invoice. A knowledge worker’s day is not eaten by tasks. It is eaten by transitions between tasks, each one demanding that a complex working state be torn down and, minutes later, rebuilt from scratch. The question that interrupted you was cheap. Paying it cost you the thread, and the thread is the expensive part.

A machine has no thread to lose

Here is the asymmetry that should reorganize how we think about interruptions. When an interruption lands on a person, it lands on a mind that was holding something, and the holding is what gets destroyed. When the same interruption lands on an agent, it lands on nothing. The agent was not in the middle of a delicate train of thought it will spend twenty minutes rebuilding. It picks up the question, answers it, and returns to whatever it was doing with no reconstruction cost at all, because for a machine the working state is written down, not balanced precariously in attention.

That is not a small efficiency. It is the whole reason the machine should be standing in front of you. The point of putting an agent between you and the interrupting world is not that the agent is smarter than you at the question. Usually it is not. The point is that the agent pays the context-switch cost that you cannot afford and it can. It eats the interruption for free.

So the question to ask of every ping, every status request, every where are we on this, is not can a person answer this. Of course a person can. The question is who pays the switch. If a person answers, the answer costs two minutes and the thread costs twenty. If an agent answers, the answer costs two minutes and the thread costs nothing, because the agent had no thread and the person it was protecting never broke focus. The same work happens. The expensive side effect does not.

This reframes a lot of what we have been calling automation. The win is not that the agent does the task instead of you. For many of these small interruptions, the task was never the point. The win is that the interruption gets absorbed by the one worker in the building who is structurally immune to it. You route the question to the thing that can take the hit and lose nothing, and you keep the thread that only you can rebuild.

It also tells you which work to point an agent at first, and the answer is counterintuitive. Not the hard, deep, high-value tasks, the ones that need a human in flow. Point it at the swarm of small, shallow, constant interruptions, the ones that are individually trivial and collectively devastating, because each one is a thread-killer for a person and a non-event for a machine. The agent’s job is to stand at the door, take every two-minute question, and make sure none of them ever reaches the person who would have paid twenty minutes for it.

The deepest work a company does happens in long, unbroken stretches of one mind staying with one problem. Every interruption is an attack on those stretches, and historically the only defense was to ask people to ignore each other, which they cannot. Now there is a real defense. Put a machine in front of the thread. Let it absorb the switching cost that was never yours to pay. The interruptions will keep coming. They just will not land on you.