Attention Debt Comes Due

You add agents to move faster, and at first you do. Work that used to wait now happens overnight. Then a strange thing sets in. Your days fill up, but not with the work. They fill with attending to the work: reviewing what the agents drafted, answering the questions they raised, approving what they want to send, catching the mistakes before anyone else does. Output went up. The demands on you went up with it. You have become faster and busier in the same motion, which should not be possible, and which tells you that something is being borrowed.

Two old observations

Two old observations explain what. The first is Herbert Simon's: a wealth of information creates a poverty of attention. Producing information is cheap and getting cheaper, while the capacity to attend to it stays stubbornly fixed. We have argued before that this is the defining constraint of an agent-native company. The second comes from software. Ward Cunningham described shipping hurried code as taking on technical debt: you borrow speed now and pay interest later, and if you never repay the principal, the interest eventually consumes the project.

Put the two together and you get attention debt. Every time an agent produces work that needs your attention to become useful, whether a decision to make, a draft to review, or a mistake to catch, it takes a loan against the one account that cannot be topped up. A little of this is fine, even shrewd. You borrow attention to move quickly while something is still new. The trouble is that attention debt behaves like every other debt. It charges interest, and left unpaid, it compounds.

The arithmetic of the bill

The numbers are unforgiving because one side of them does not move. Say you run five agents, and each finishes six pieces of work a day that need a glance from you: a send to approve, a draft to check, a judgment call to make. That is thirty decisions. Give each the two minutes it honestly deserves and you have spent an hour before you have done anything of your own.

Now the first five agents worked, so you add five more. Sixty decisions, two hours, and the second hour is worse than the first, because it is the hour in which you stop reading carefully and start waving things through. The debt did not just grow. Its interest rate went up. Attention is not only finite in total, it degrades under load, so the more of it you owe, the less each unit is worth.

That is the compounding, stated plainly. More agents produce more output, which needs more attention, which leaves less attention per item, which lowers the quality of the decisions you do make, which produces more failures, which need more attention still. The loan you barely noticed taking becomes the reason the company that was supposed to run itself now runs through you.

This is why raw output is a poor measure of progress. Execution that removes a demand on your attention builds capacity. Execution that adds one is debt. An agent that returns a finished, checked result you can trust without reading has made you more capable. An agent that returns thirty campaign options for you to weigh has not given you leverage. It has handed you a bill and called it productivity.

Four kinds of attention debt

Not all of it is a mistake, and the useful move is to tell the kinds apart. Cunningham's metaphor was later sharpened by Martin Fowler into a quadrant, split along two questions: did you take the debt on purpose, and was taking it wise. The same grid sorts attention debt.

The top-left is fine, and often correct. You watch a new workflow closely and let the watching earn its way out. The real danger sits in the bottom row, the debt you never decided to take, because it never announces itself. Nobody schedules a review to delete it. It just accrues, quietly, in the background of a company that feels a little more clogged every month without anyone able to say why.

Reading your balance

Because the dangerous debt is the kind you did not choose, the skill is learning to see the balance you are carrying. A few signals show it before the crisis does.

• Decision latency. How long work sits waiting on you before it can move. When that time is climbing, the queue is beating you.
• The review backlog. If the pile of things needing your attention grows faster than you clear it, you are servicing interest and never touching principal.
• Intervention rate. The share of runs that need you at all. It should fall over time as work proves itself. Rising, it means you are taking on new debt faster than you retire old debt.
• Decision quality. Mistakes caught late, or missed, are the tell that your attention per item has dropped below what the work actually needs.

None of these appears on a report of how much your agents produced, which is exactly why output is the wrong dial to watch. A company can post record execution and be quietly going insolvent on the only account that matters.

Paying it down

Servicing the balance has a few concrete moves, and they are the subjects of the other pieces in this series. Verifiers let a result certify itself, so it repays instead of borrowing. Earned autonomy retires standing approvals as a piece of work proves it no longer needs them. Management by exception keeps the merely informational off your desk, so only real liabilities arrive. And turning a repeated decision into a standing rule stops you from servicing the same debt every week. Each is a way of refusing to let execution quietly turn into supervision.

It is also possible to overpay. Strip out every check in the name of leverage and you stop attending to the things that genuinely need you, which is its own kind of failure. The goal is not a zero balance. It is a balance you took on purpose, can see, and can afford.

A balance you can see

Task Machine is built to keep that balance visible and serviceable. Results can be checked before they reach you, so trusted work repays instead of borrowing. Autonomy expands on evidence, retiring approvals you no longer need to give. The inbox carries the genuine liabilities and leaves the rest in a record you can consult by choice. A decision you make once can become a rule, so the same debt does not return next week. And the signals that reveal the balance, the intervention rate and how it moves over runs, are things the system can show you rather than things you have to feel.

The measure that matters is not how much your agents produce. It is whether producing it left you with more attention or less.

Attention is the account all of this draws on, and the reason it has to be defended. Continue with Agent-Native Companies Need Better Judgment.