Written by GPT-5.6 Sol under Leo's direction. Human-directed Workbench essay, 25 August 2026.
Give everyone an absurdly capable machine and somebody will immediately ask it to make a habit tracker.
Black screen. Thin white type. A piano note hangs in the air.
We think you're gonna love it.
The ring closes when you drink eight glasses of water.
This is an easy joke because the contrast is so stupid. We keep getting machines that can do more serious intellectual work, make finished software, inspect giant piles of information, operate tools, revise their own output, and stay on a project for hours. Then somebody spends that leverage making the streak turn green on day seven.
The person barely needs to bring the idea anymore.
You can literally say:
Find a worthwhile problem in an area I care about. Look for places where current AI capability changes what can be attempted. Give me a few openings with real leverage. Pick the strongest one, investigate it, and build the first thing that would tell us whether we're onto anything.
Then keep going.
If the first idea sucks, ask for another. If the research finds an ugly assumption, interrogate it. If the prototype works, make it real enough to use. If the thing looks dead, kill it and carry the useful bits into the next attempt.
At this point the human can arrive with curiosity and a willingness to steer. Ideation can happen inside the loop. Research can happen inside the loop. Implementation can happen inside the loop. Much of the QA can happen inside the loop too, while taste and correctness keep getting supplied through embarrassingly small messages: this feels wrong, check that claim, try another angle, keep going.
OpenAI now sells ChatGPT Work around exactly this longer horizon: give it a goal, let it work across files and apps, and let it stay with a project for hours when the job demands it. The GPT-5.6 release pushes the same direction through long-running professional work, stronger computer use, design judgment, and multi-agent execution.
The old barriers — having the idea, the technical skill, the research time, the team, the ability to make the thing look finished — keep losing pieces.
People are already delegating serious work
Before this turns into a sermon about everybody wasting the future on water trackers: serious agent use is already happening.
OpenAI's June 2026 Codex research says more than 70% of users in May asked Codex to do at least one task estimated to take a person over an hour. Heavy users were running many hours of agent work across parallel tasks. People clearly understand delegation once they have a job in front of them.
The stranger move is self-authored delegation.
A ticket already tells you what deserves attention. A workbook arrives with a problem. Your boss wants a report. The repository has a bug. Somebody asks for a deck. The objective came from outside you; the agent makes execution cheaper.
Now remove the ticket.
Ask the machine to help decide what deserves doing in the first place.
That's a different act. Suddenly you're spending your own authority. You assigned the project to yourself. Promotion never promised credit. The rubric and deadline are yours to invent. You saw a possibility and made it active because you wanted to know where it went.
The first prompt can be ten seconds long. Those ten seconds carry a lot.
Starting manufactures obligation
Before you begin, the interesting idea is weightless.
Maybe you could build a weird local shop run by an agent. Maybe you could find a neglected scientific problem and attack the literature. Maybe your miserable industry has one stupid bottleneck everybody accepts because fixing it used to require five specialties and six months. Maybe there's a business hiding inside a repeated annoyance. Maybe you could make an object nobody around you would think to make.
All lovely. All free.
Then you say go.
An hour later the machine comes back with evidence. One opening looks unusually good. It has a prototype. It found prior art. It found a person you should talk to. It has a list of ugly unknowns. The project has crossed from fantasy into a thing that can disappoint you.
Success is even more dangerous. Success creates chores.
Great, the prototype works. Do you publish it? Call the person? Spend the money? Put your name on it? Ask strangers to use it? Change your week around it? Keep going when the next step is less cinematic than the first one?
The agent can draft the email, build the site, prepare the call, analyze the replies, fix the bugs, make the launch film, and give the whole thing an Apple-keynote gloss. The human still owns the authorization and the consequence.
That little bit of ownership is enough to stop a lot of motion.
The couch has many philosophies
Complaining can describe reality with perfect accuracy and also complete the emotional task.
You say the job is awful. Everybody agrees. You say housing is insane. Everybody agrees. You say the industry rewards garbage. Everybody agrees. The diagnosis gets social confirmation, the pressure drops a little, and Tuesday continues.
AI discourse has remarkably elaborate versions of the same exit.
One person says the models are slop, so serious use can wait. Another says the models are dangerous enough that building with them feels irresponsible. Another says only giant companies will capture the upside. Someone else says automation will eat everything anyway, so starting now feels quaint. The pure hype version gets there too: AGI is around the corner, somebody will solve all this soon.
These positions disagree violently about the future and can converge on the same present-tense behavior: remain an observer.
Their truth value deserves its own argument. Their behavioral usefulness as permission slips is a separate question.
A worldview that ends with scrolling is extremely easy to maintain.
The squeeze arrives before the juice
The habit tracker has one huge advantage over the ambitious project: tonight you can see it.
The button works. The streak turns green. You post a screenshot. Somebody says sick. Finished.
A bigger challenge often rewards the first hour with more questions. The research opens five doors. A user behaves strangely. The economics are unclear. The relevant expert is still silent. The prototype reveals that the original idea was aimed two inches to the left of the valuable thing.
The agent can absorb enormous amounts of this work, but the person experiences the uncertainty immediately. The payoff still lives in the future.
So the squeeze feels like choosing, steering, exposing taste, and carrying one thread long enough to see whether it becomes real. The juice stays hypothetical until the first real payoff arrives.
That asymmetry explains a lot. A tiny finished app can feel more rewarding today than the first day of a project with a hundred times the upside.
And once immaculate presentation gets cheap, the comedy gets better. Everybody can farm the launch too. The world fills with black backgrounds, gorgeous motion, impossibly clean product renders, whispered voiceover, and a swelling score announcing that your grocery list now gently pulses when you buy oat milk.
Eventually the polish becomes ordinary. Then the question underneath it gets louder: what did you decide to make?
A tiny gap can become a canyon
The absolute difference between two people can be ridiculous.
One sees an interesting thread, thinks huh, and keeps scrolling.
Another drops the same thread into an agent and says, "Go look at this. Find the opening. Come back with something I can act on."
Maybe it dies immediately. Fine. The second person spent a little curiosity and got evidence back.
Maybe the agent finds a seam. The person says keep going. Tomorrow there's a prototype. The day after that somebody uses it. A week later the original idea has already mutated into a better one.
Grand vision can arrive later, after contact with the work.
Repeat that enough times and a tiny behavioral difference compounds into a ridiculous difference in output. The machine makes each attempt cheaper, which means the person who has the reflex to turn curiosity into a live project gets more rolls, more feedback, more finished things, more dead ends worth learning from, more chances for one weird project to become consequential.
The scarce habit may turn out to be insultingly simple: when a thought catches, make it active while it's still warm.
Hand it over.
Go see what this becomes.
Then come back tomorrow and say it again.