The question that is not interesting
Here is a made-up but representative number, the kind of split you'd get from timing a real task on your own desk: a task took ninety minutes last month and takes ten now. That is not eighty minutes saved. It is eighty minutes saved minus whatever it costs you to be sure the ten-minute version is actually right — five minutes of checking, and the trade is extraordinary; forty minutes of checking because the output is large and plausible-looking, and you saved forty while quietly taking on a new risk you have not priced yet.
Whether a model can write a function is not the interesting question. It can, most of the time, and arguing about the exceptions has consumed three years of everyone's attention. The interesting question is what happens to the rest of your day once the doing stops being the slow part. That is a question about your working week rather than about the model, and the answer is not "you get more done." It is stranger than that, and it is uncomfortable.
What actually changed: the agent has hands
Autocomplete suggested. You accepted or you didn't, and nothing happened to your files until you said so.
An agent with tool access reads your actual files, runs your actual commands, and changes things before you have looked. That is not a faster version of the previous tool. It is a different relationship, and every consequence below follows from it.
A suggestion is a proposal you evaluate at leisure. An action is a fact you discover. Everything that is hard about working this way comes from that one sentence, and everything that works comes from designing around it.
Three things change immediately:
Your unit of work gets bigger. You stop asking for a function and start asking for an outcome — make the export handle the missing-country case rather than write a country parser. The instruction gets shorter and the surface it touches gets wider.
Your attention moves to the end. You used to spend it while producing. Now you spend it after, on output you did not watch being made. This is the shift people feel as disorienting and misdiagnose as "I don't trust it."
The cost of asking collapses. Trying a second approach used to cost an afternoon; now it costs a sentence. That sounds purely good and is not — see the section on what gets worse.
What did not change, and gets ignored
Knowing what you actually want. The failure mode of a capable agent is not incompetence, it is fluent completion of the wrong task. A vague instruction used to produce visible confusion. Now it produces a confident, finished, wrong result. The specification burden went up.
Responsibility. The output ships under your name. No model, and no vendor's terms of service, absorbs that. This is not a moral point; it is a practical one about who has to fix it in March.
And the boring parts of your business. An agent does not know that the invoice must go out before the 5th because your customer's finance team runs a batch on the 6th. That knowledge lives in your head, and until you write it down it is not delegable. Most of what makes a small business work is exactly this kind of undocumented judgement.
The bottleneck moved, and this is the whole argument
Generation got cheap. Checking did not. Every operating decision on this site hangs on that gap.
Generation got cheap. Checking did not.
That asymmetry is the entire thesis of this site. It sounds obvious stated flatly and almost nobody has restructured anything around it.
Work through that same arithmetic on your own week, not just the illustrative ninety-to-ten example above. Every task you delegate has its own version of that five-minutes-versus-forty split, and the honest number is rarely the one you'd guess without timing it.
And if you skip the checking, you did not save eighty minutes. You deferred an unknown cost to an unknown date. That is the trade almost everyone is currently making without writing it down.
Not a dramatic wrong answer. A plausible one. Three composite patterns, each seen enough times across enough different tasks to be worth naming even without one specific incident attached: a report that looks right and uses last quarter's figures. A cleanup that worked and silently dropped forty rows. A refactor that passes every test because it also updated the tests.
Confident and wrong is the expensive failure mode, and it is the one that scales with capability rather than away from it.
Why this is an economics problem, not a tooling one
Ronald Coase asked, in a 1937 paper called "The Nature of the Firm," why firms exist at all rather than everything being bought on the open market. His answer: coordinating work inside a hierarchy is sometimes cheaper than negotiating it every time. Firms grow until organising one more thing internally costs more than buying it outside.
Agents do not repeal that. They change which cost is binding. Coordination — the meetings, the handoffs, the explaining — collapses toward nothing when the other party is an agent that reads your files. What replaces it as the limit is verification: how much output you can confirm is correct before it compounds.
So the size of what one person can run is now set by their checking capacity, not their effort. That is a different constraint with different levers, and the levers are learnable — which is fortunate, because effort was never very learnable.
The lineage matters here and is credited deliberately, as an analogy rather than a rigorous mapping: Coase and Williamson on transaction costs, Herbert Simon on bounded attention, and Michael Jensen and William Meckling on what it costs to monitor an agent whose interests you cannot fully observe. None of them were writing about software. But the structure each one described — a coordination cost, an attention limit, a monitoring cost — is the same shape as the checking constraint this piece argues is now binding, which is why the borrowed language is more useful here than a fresh vocabulary invented from scratch would be.
The four habits that make it work
Not tips. These are the four that survive contact with real work, and each one exists to make checking cheap.
The first task to delegate is not the most annoying one. It is the one where you can tell fastest whether it worked. A list to deduplicate has a countable answer. A strategy memo does not. Build the habit where the feedback is immediate, and you will have it before you need it somewhere the feedback is slow.
One change at a time, in a place you can undo. Not because the agent is unreliable — because a large change is a large thing to check, and an unreviewable diff gets approved rather than read. Work in a copy, keep the original, and a bad run costs one repeat.
The agent will tell you what it did. That description is an instrument; the diff is the ground. When they disagree, the diff is right — and the disagreement is where the interesting bugs live. This is the single habit with the highest return, and the one most often skipped because the summary is so much easier to read.
Do not ask it to clean the data and restructure the file and fix the naming. Not because it cannot — because when something is wrong you will not know which of the three did it. Bundled changes are cheap to make and expensive to diagnose.
Ask for things whose correctness you can establish quickly, in steps small enough to reverse, and read what actually changed. Everything else is elaboration.
Where this genuinely does not work yet
Stated plainly, because a page that only lists advantages is marketing.
Work where being wrong is expensive to discover. Anything that leaves your control and takes weeks to surface — a pricing change, a filing, something a customer relies on quietly. The check has to happen before it ships, and if you cannot construct that check, the task is not ready to delegate.
Work where recovery is hard. Producing at speed is easy; unwinding at speed is not. A cascading error across three systems that each accepted it takes far longer to reverse than it took to create. If your recovery cost grows faster than your output, the leverage inverts — you have built a liability engine rather than a productivity one.
Work that depends on what you have not written down. The undocumented judgement. Every step of making it explicit is real work, and it is the actual cost of this transition that nobody quotes.
The number that tells you whether any of this is true
Time one task. Actually time it — before and after. Then time the checking separately, because that is the number everyone forgets and it is the one that decides whether you gained anything.
Two honest outcomes:
The task is faster and checking is cheap. You have found something real. Do it again next week with a different task and start keeping a list.
The task is faster and checking eats the gain. Also useful, and more informative. Either the task was the wrong one to start with, or you have found where your verification tooling needs to exist — which is a specific, solvable problem rather than a vague dissatisfaction.
Anyone who cannot tell you which of those two they are in is guessing about their own business.
A Zero-Employee Organization (ZEO) is not a company with nobody in it. It is a business that has asked that question of every task it runs, has the numbers, and has stopped needing to add people to add output. The arithmetic above, repeated across a working week, is the whole method.
The shortest useful version: clean a messy customer list in about twenty minutes, and prove nothing was lost. It is where the checking habit is cheapest to build.
