The Work Release Desk

Everyone I talk to is producing more work than they have ever produced in their lives. Operators who never wrote a line of code are shipping working software. Leaders who used to decide on instinct now walk into the room with ten pages of research behind every choice, because the research costs twenty minutes instead of two days. The scope of what one person can attempt has exploded, and the depth behind each attempt has exploded with it.

And somewhere in the second half of almost every one of those conversations, the same person lowers their voice and says a version of the same thing. I cannot keep up with my own output. I am drowning in work that I chose.

I don’t think that is a personal failing. I think it is physics. And I think the fix is structural, not motivational. Stay with me through a little factory theory, because it ends in practical advice.

The math of why you are tired

There is a sixty-year-old result in queueing theory called Little’s Law. In its most useful form it says: the time it takes to finish anything equals the amount of work you have in flight, divided by the rate at which you actually complete things. Cycle time equals work in progress over throughput. Notice what is not in that equation: effort, talent, tools, hours. If you double what is in flight without doubling your completion rate, everything takes twice as long. Not feels longer. Takes longer.

Queueing theory adds a second, nastier result. As a system approaches full utilization, waiting time does not grow gradually. It stays flat, and then it goes vertical. A highway at 70% capacity flows. At 95% it is a parking lot. The curve has a knee, and the knee arrives suddenly.

Now look at what AI actually changed. It did not make you faster at finishing. It made starting nearly free. Every idea can become a running prototype by lunch. Every question can become a research report before the meeting. The arrival rate of work into your life went vertical, and the capacity of the finishing step barely moved. In a 2024 Upwork Research Institute study, 77% of workers using AI said it has added to their workload. Faros AI’s telemetry across 22,000 developers found that in teams with heavy AI adoption, task throughput rose 33.7% while deployments per week fell 11.7%. More starting, less finishing.

The difference between those two numbers is inventory. Donald Reinertsen made the point years ago that product development inventory is physically and financially invisible, and the inventory of the AI era is more invisible still, because it is parked in your head. Nobody trips over a pallet of unreviewed decisions. You just carry them, and they are heavy.

The slowest station in the factory is you

The Theory of Constraints, Eliyahu Goldratt’s contribution in The Goal, says a system moves at the speed of its slowest step, and that improving anything other than the slowest step is an illusion. His hiking troop moves at the pace of Herbie, the slowest kid, no matter how fast everyone else walks.

In AI-assisted work, it is worth naming the slow step precisely. It is not typing. It is not generating. It is not even reading. It is the trust decision: the moment you say, this is good enough, and I am accountable for it. A recent study of 802 developers carried a title that summarizes the entire era: AI Writes Faster Than Humans Can Review. Output per person doubled. The review load per reviewer doubled right behind it.

Why is that one step so expensive? Three reasons compound.

First, the almost-right problem. In Stack Overflow’s 2025 survey of 49,000 developers, the top frustration with AI, named by 66%, was solutions that are almost right but not quite. Obviously wrong output is cheap to reject. Output that is 95% right is expensive, because finding the last 5% means reconstructing the reasoning of an author who does not exist.

Second, the verification asymmetry. Generating a plausible artifact is now nearly free. Verifying one never got cheaper, because verification means holding the whole thing in your head and hunting for the case that breaks it. The Anthropic Economic Index put a price on this: adjust for the human time spent validating AI output, and the implied productivity gains drop from 1.8 points to roughly 1.0 to 1.2. A third to a half of the gain is paid right back at the checkpoint.

Third, unreviewed output does not politely wait. It transfers its cost to whoever receives it. The researchers who coined the term workslop found 40% of US desk workers had received AI-generated work that looked finished but was not, and that each instance cost the recipient nearly two hours. Someone saved twenty minutes and spent two hours of a colleague’s life.

And here is where the factory metaphor breaks in a way that matters. A milling machine does not slow down because the last part was borderline. Judgment does. This station depletes with use. Simon Willison, one of the best practitioners alive, described running four agents in parallel and being wiped out by 11 AM. Addy Osmani caps himself at three or four threads and names the ambient anxiety tax of every thread he cannot fully watch. His summary is the one to remember: your cognitive bandwidth does not parallelize.

The work release desk

Factories that run on the Theory of Constraints follow a rule that sounds backwards until it saves you: you do not release work into the plant based on the availability of the first station. You release it based on the tempo of the bottleneck.

Our first station is generation, and it is now free and infinite. Releasing work based on its availability means starting something every time we have an idea. That is the entire failure mode of this era in one sentence.

The alternative is what I have started calling a work release desk: a deliberate gate that decides what gets admitted into your system and what actually goes out the door. I stress tested this idea against three bodies of thinking, the operations science, the attention research, and the lived practice of people who run AI agents all day. Here is the version that survived.

Give the desk a number. Three or four slots for work that requires your judgment, written somewhere you can see. A slot opens only when something ships or gets killed, never because something is merely in progress. Finishing is the only currency that buys a new start. Cal Newport caps his active list near three. Osmani’s agent ceiling is three or four. The convergence is not a coincidence. That is roughly how many open judgment loops a human can hold.

Agents wait. You do not get pinged. Work in progress accumulates in a queue that you visit, not a stream that visits you. Serve it in two or three review windows a day, batched, with one context loaded at a time. Knowledge workers already get interrupted about every two minutes; routing agent output into that storm is self-sabotage. Batching your review is not laziness. It is protecting the machine the whole factory depends on.

Match decisions to your sharp hours. The trust decisions with the biggest blast radius go into the hours when your discrimination is sharpest, usually morning. Mechanical, easily verified work goes into the fog hours. The 11 AM wall is a scheduling input, not a moral failing.

Only gates that can say no add capacity. Tests, type checks, CI gates, evals: these genuinely take load off your judgment, because they can reject work without you. An AI reviewer that approves things is different. Approval without accountability is laundering, not review. Faros found 31% more pull requests merging with no review at all. That is what it looks like when a system quietly routes around its own bottleneck instead of managing it.

Shape the work smaller before it starts. Queue length is driven by volume, but also by variability, and reviewing AI output is wildly variable: ninety seconds if it is cleanly right or cleanly wrong, ninety minutes if it is almost right. Smaller work items, tighter specs, and an acceptance condition written down before the agent starts will shrink your queue without you refusing anything.

If you cannot judge the output, judge the exposure. This one is for the non-technical builders, and I meet more of you every month. You can ship now, but you cannot yet look at code and know if it is good, and no slogan about discernment fixes that. So judge what you are already qualified to judge, before anything is built: who could lose money if this is wrong, whose data is inside it, and can I undo it in ten minutes. Low exposure, ship freely and learn. Anything above that does not ship without a named verifier who is not you and not an AI.

What “separate” buys you

The desk only works if it is separate in two ways.

The part that is not about throughput

The overwhelm is not only a speed problem, and the research here is unusually actionable.

Unfinished work pulls at you even when you are not working on it. But a 2011 study by Masicampo and Baumeister found something remarkable: you do not have to finish a task to release its grip. You have to make a specific plan for it. Writing down the next physical action and where the materials live eliminated the measured interference of unfinished goals on unrelated work. So end every review window, and every day, by closing each open thread in writing: next action, where it lives. Unwritten threads charge rent overnight.

And treat recovery as the buffer in front of the bottleneck, not as a reward for clearing the queue. The queue never clears. Oliver Burkeman calls this the efficiency trap: becoming more efficient just makes you more rushed, because freed capacity is consumed by new demands the day it appears. So decide in advance where the freed capacity goes, and put some of it on the calendar before the slots fill up. A buffer is not indulgence. It exists to absorb variability, and you are the one component in this system that breaks quietly, looks fine in every throughput metric, and does not recover on a schedule.

The factory can now run as fast as you dare. You are the only station in it that cannot be scaled, and the only one the factory exists to serve. So put a desk in front of your judgment, give it a number, and let it say no on your behalf.

The ideas will keep. The queue can wait. That is what queues are for.