Why we work from queues.

What an inbox is, seen as a system

Picture a bathtub. The water in it is everything you have been asked and have not yet dealt with. The tap is other people sending you things. The drain is you, answering.

Three facts about this bathtub make it a poor way to run work.

You do not hold the tap. Anyone can add to your inbox, at any moment, at almost no cost to themselves. The cost lands on you.

Nobody can see the level. The person about to send you the fortieth thing today cannot see the other thirty-nine. Often you cannot see it clearly yourself: an inbox mixes what needs a decision with what is merely news, in the order it happened to arrive.

There is no limit. Nothing stops the water rising, and nothing tells the people at the tap to slow down when it does.

Nobody chose this as the way to run an organisation's work. Email arrived in offices as a better way to send a message, and became the place where tasks are handed out, tracked and remembered. Researchers watching office workers in 1996 already found them using the inbox as a to-do list and an archive, jobs it was never built for.

What goes wrong, and why

The wait grows much faster than the load. This is mathematics, not opinion. For work that arrives unevenly, a person who is busy half the time has on average one thing in hand or waiting. At 80 percent busy it is four. At 90 percent, nine. At 95 percent, nineteen. Short waits need spare capacity. And if more arrives than can be done, the pile does not settle at a large size. It grows without end.

People do not simply fall behind. They lower the standard. When more arrives than a person can do properly, they answer faster and shorter. A study of a bank's service centre found that a temporary surge, met by overtime and by doing each job less thoroughly, produced a lasting fall in the quality of service: the lower standard became normal. Related work describes the trap that follows: under pressure, time goes to keeping up and none to fixing how the work arrives, so the pressure never lifts.

Interruption costs strain. In a controlled experiment, people interrupted about every two minutes while answering mail finished slightly faster, and no worse, than people left alone. They reported markedly more stress, frustration and effort after only twenty minutes. People can absorb interruption for a while. They pay for it.

Work that is left gets lost. Observation of office workers found that more than half of their stretches of work were interrupted; that about a quarter of interrupted work was not taken up again that day; and that when it was, about 25 minutes had passed, spent on other things. The danger is less the minutes than forgetting where you were.

The same is true for an AI agent, more sharply. An agent's conversation is its inbox. Published studies find that language models attend poorly to what sits in the middle of a long conversation, and become less reliable as a task is spread over many turns. And an agent's conversation ends: the next session starts knowing nothing. Work that lives only in a chat is the work most likely to be dropped.

What a pulled queue changes

A pulled queue is a list of work that waits until someone with room for it takes the next item. The idea comes from Toyota's factories and has since been carried into software, laboratories and offices. Its core is one rule: work enters only when there is room for it.

That rule changes each of the three facts.

The worker holds the tap. Things still arrive, but they arrive in the queue, not in the middle of what you are doing. You take the next one when you finish the last. At a genomics institute that made this change, a group carrying twice the work it could finish cut its work in progress by more than half and finished several times more.

The level can be seen. Every item is in one place, with a count, a kind and a score. You can see what is waiting, how long, and for whom. So can the people who would add to it.

There can be a limit. A queue can say how much fits in a day and hold the rest for tomorrow, in the open, with a reason.

Four more things follow.

• Order is decided once, at the door. In an inbox the order is the order of arrival, re-decided every time you look. In a queue each item is scored when it enters. You do the top one.
• A queue can be finished. A queue for today, sized to today, can reach zero. That is the point of the design.
• The item carries what the next person needs. A task in a queue says what done means and what the worker needs to start. A forwarded chain of mail says neither.
• One shared queue is faster than many private ones. When several people or agents can do the same kind of work, one queue they all draw from wastes less time than a queue each.

In Infinity, as designed: everything that needs a person arrives in one queue, as a question with choices, a proposal to approve, a review, a task, or a notice of something already done. Agents score each item. The day is split into quick items, items an agent is stopped on, and deep work sized to the hours available. Each list can reach zero. Agents work the same way: a request to an agent is a task addressed to it, claimed so that no two do it, and reported on when done.

Where it does not help

A queue is a tool with costs, and there are things it does badly.

• New, tangled problems need talk. When nobody yet knows what the question is, two people talking is faster than any exchange of items. Infinity keeps rooms for that, attached to the thing being discussed: talk in the room, and put in the queue only what needs someone to act.
• It does nothing about the interruptions you cause yourself. About half of all interruptions observed in offices were the worker's own.
• A queue between two teams can become a wall. Ticket systems can let teams stop talking, lose context at each hand-off, and reward closing an item over solving the problem.
• A system with no slack breaks under shock. A person's queue sized to exactly a full day has the same weakness as a factory with no buffer.
• A limit nobody enforces is decoration. The hardest part, in every account of pull, is refusing to push one favoured item past the rest.
• It has to remind as well as the inbox did. People keep things in their inbox because they stay in sight. A queue that hides work will lose to the inbox, and deserve to.
• It costs something to write a good item. That cost is the price of the next person not having to ask.

How sure we are

No one has run the experiment that would prove a pulled queue beats an inbox for knowledge work. What exists is mathematics that is certain under its conditions, a method proven in manufacturing, one strong case study in a research institute, small experiments on stress, and a description of the inbox as a system that holds up. For AI agents the published research supports the diagnosis, and nobody has yet tested the cure. We think the case is strong, and we hold it as a case. The way to find out is for a queue to measure itself: what arrives, what is finished, and how old the oldest item is. That is proposed and not yet built.

Sources

• Queue lengths: the standard single-server queueing result, average number in the system equal to load divided by one minus load, under its stated conditions.
• Email used as a to-do list and archive: Whittaker and Sidner, 1996, a study of office workers.
• The bank's service centre and the lasting fall in standard: Oliva and Sterman, 2001.
• The capability trap: Repenning and Sterman, 2001.
• Interruption and stress: Mark, Gudith and Klocke, 2008, a laboratory experiment with 48 participants.
• Interrupted work and resumption: Mark, Gonzalez and Harris, 2005, observation of 24 office workers.
• Pull at a genomics institute: Dodge, Kieffer and Repenning, 2018, a case study of the Broad Institute.
• Language models and long conversations: Liu and others, 2023; Laban and others, 2025.
• Just-in-time and pull: Toyota's own description of its production system.