Twenty years ago, a few weeks into my job at Google, I wrote a post about 20% time. For anyone who never ran into it: 20% time was Google’s policy of letting engineers spend a day a week, a fifth of their time, on a project of their own choosing. The argument in that post was that you couldn’t just copy it. 20% time worked because of the environment around it, not because someone wrote “20%” in a handbook. It was, as I put it then, “a result of an environment and philosophy to development more than a cause.”
I still think that’s right. The part I keep coming back to in 2026 is the optimism underneath it. 20% time was a hopeful bet: give people room to wander, and trust that good things fall out. Almost nobody runs it as a policy anymore, Google included. But that instinct is back across the industry, this time as a side effect of AI rather than anything in a handbook. That’s the new 20% time: the promise is back, not the room itself. The old version was paid in hours, so you could set a day aside. This one is paid in attention, and there’s no setting that aside. The title’s “minus the time” is the catch. The promise comes back without the one thing that made it real, which means it starts as 120% time. The question underneath, the one I can’t shake, is “Who benefits?”
A day a week was never the point
The thing people remember about 20% time is the number itself. That part was almost incidental. In the 2005 post I laid out several things that made it function, and only one was about time. It all came together to describe an environment that trusted engineers to point themselves at something useful. The 20% was just the permission slip. The culture was the actual thing.
That Google is mostly gone, which is what happens when a company goes from a few thousand people to nearly two hundred thousand. 20% time got quietly walked back along the way. Around 2011, Larry Page returned as CEO with a “more wood behind fewer arrows” push, and Google Labs was shut down. By 2013 engineers were calling it “120% time”: still allowed in theory, but you needed your whole week to hit your targets, so the side project landed on top of everything else. It died because the company got good at measuring output, and once that’s the baseline, any hour you spend wandering is an hour you’re behind. Room to explore only survives when nobody is measuring it too closely. Hold that thought.
What’s actually scarce
Quick definition first, because “side project” was slippery even inside Google. For me it isn’t something unrelated. It’s the speculative end of the job itself: the things that don’t fit our immediate plans but probably should exist. Still my work, just the part that isn’t on the roadmap yet.
In 2005 a 20% project was a block of time. You took a day, or stole an afternoon here and there, and went deep on something that wasn’t your day-to-day. The unit was hours. The scarce thing was a stretch of uninterrupted concentration long enough to hold a hard problem in your head.
That part of the job doesn’t work the way it did. Coding used to feel like cutting and welding metal, every change a small fight with the material. With an agent it feels more like clay: you push the shape around, reshaping costs almost nothing, and the clay never decides what you’re making. So I start the speculative thing while the main work is still running. I hand an agent a task on what I’m supposed to ship, and while it grinds away I chase the thing I’m curious about in another window. A few minutes later the first one wants a decision, so I go back. The exploration never gets a clean day. It gets the gaps that open every time the main work doesn’t need me. It’s 20% time without the calendar, and nobody has to approve it.
But the constraint moved. Hours stop being the bottleneck once the agent does the typing. Attention is. The effortful kind: the part of you that holds context across several threads and decides whether what the agent just handed back is actually right.
Running a handful of agents and switching between them is its own kind of tired. Harvard Business Review has started calling it “AI brain fry”: the strain of supervising systems that move faster than you can think. It isn’t burnout. For me, burnout is feeling responsible for something you can’t affect, shoving at an immovable object. Brain fry is almost the reverse. You have all the help you could ask for, a row of agents that will do whatever you tell them, and the thing that gives out is your own capacity to hold context and decide. We spend all our attention on what the tools can do. We’ve barely started on what running them does to the person doing it.
Who collects the dividend?
AI reaches well past software developers, and not as a gift.
Step back and the through-line is almost hopeful: AI is handing us more room to explore than we’ve had in years. But the same technology reads as a catastrophe one field over. Ask an illustrator, a voice actor, or a screenwriter how AI is going and you won’t hear “more room to explore.” You’ll hear that the thing they spent a life learning to do was scraped to train the tool, and now the tool undercuts their rate and, in a growing number of jobs, replaces them. The honest difference is who the tool is pointed at. For me it speeds up the work. For them it’s aimed at doing the work without them. And the agent buying me that room was trained on theirs: my dividend may be built on what was taken from them. Everything in this post about more room to explore happens to people like me. The 20% optimism was never on the table for them. From where they’re standing this isn’t room to explore. It’s work being taken, often by the same tool that’s buying me mine.
Even where the optimism is real, I don’t know who ends up benefiting. That’s the part the cheerful version skips. And I should be honest about what the gaps actually are. What mattered in 2005 was never the day. It was protected time: room that was sanctioned, not stolen. The gaps between agent runs are the opposite. Nobody set this time aside. It’s the same 120% time that killed the original, the side project piled on the real one.
20% time was a luxury good, slack paid for by an extraordinarily profitable business, and when that business wanted it back, it took it back by learning to measure. 20% time died of measurement; its replacement is born under the microscope. I said room to explore only survives when nobody is measuring it too closely, and a token bill measures itself. Maybe someone else is paying that bill for now, though nobody knows how long that lasts. The other cost is yours no matter who pays: attention is finite, and no one is cutting you slack there.
20% time, at its best, was optimism made concrete: a company handing people room and getting something good back. AI is the first real chance in years to make that bet again, at a scale no handbook could reach. So the real question isn’t whether AI saves labor. At least for me, it does. And that saved labor is a dividend. The question is the one I opened with: who benefits?
Maybe those benefits come back to us as room to explore and make things. Maybe they pool with the people who can already afford to point a swarm of agents at whatever they want. Or maybe they just get turned back into pressure on workers. Fifty years of productivity gains have accrued to the top, not those doing the work. I want this one to break the pattern. I can’t, with a straight face, say it will.