The host opened the segment with a joke against himself. People keep talking about a three- or four-day working week, he said; he has discovered "a nine-and-10-day work week." He was introducing a Wall Street Journal article, which he summarized as confirming what the panel had been feeling: productivity gains from AI agents are creating more work for humans, not less.
In his account of the piece, agents produce more output, and that output needs more review, more decisions, more direction and more human judgment per unit of time. A founder who used to manage five tasks now manages fifty agent outputs. "The bottleneck is shifted from execution to judgment," he said. The work is not disappearing; it is changing character.
An agent, in the sense used throughout the episode, is an AI system that takes a goal, plans steps and uses tools to carry them out — drafting, searching, coding, running research — rather than answering a single question and stopping. Several of the panelists run many at once.
Earlier in the show the host had described the underlying problem as an "impedance mismatch": very fast, increasingly capable tools meeting institutions, companies and individuals that move at a linear pace. That leaves two options, as he put it. Hand the work over to the AI completely with an objective function — make me maximally profitable, run my government more efficiently — or try to sit in the middle. When the human is standing at that interface, he said, it breaks very quickly.
Jevons paradox, applied to attention
Salim Ismail named the mechanism. This is Jevons paradox for human cognition, he said, invoking the nineteenth-century observation that making a resource cheaper to use can increase, rather than reduce, how much of it gets consumed. "We thought AI would reduce workflow. In fact, it increased the amount of work that is worth attempting."
He illustrated it with his own books. The first, written with the host, took three years — "three years of hell." The second took two and a half years, also hell. The third took six months, with "a lot of joy, but damn, overload on the cognitive workload." The calendar shrank by a factor of six; what it cost him did not shrink with it.
His description of the result was structural. If ten agents report to a founder, he said, "we've reinvented middle management. It's inside your own brain." The founder is now doing the work a management layer used to do — reviewing, prioritizing, approving — except the reports never sleep and the queue never empties. Faced with that volume, in his phrase, judgment and attention become absolutely paramount.
And the arrangement cannot hold, he argued: "you can't have machines operating in machine speed and requiring a human approval on that." He said he was running into it himself in that day's work.
What Ismail says has to be designed next
Ismail's proposed remedy is not a better model but better delegation. The next breakthroughs, he said, need to be in delegation, permissions and escalations — and he described designing exactly that in a pilot program taking a group of companies through the process.
The design question he puts to those companies is a set of boundaries: what the machines may decide autonomously, what the humans want to manage later, what gets audited after the fact, and "what genuinely needs a human." A permission boundary fixes what an agent can do without asking. An escalation threshold decides which cases get pushed up to a person. A retrospective audit accepts that the action already happened and checks it afterwards, which is the only way a review can run slower than the work it reviews.
He described this as requiring a new threshold of escalation rules and governance inside firms, and he framed the alternative bluntly: otherwise the future is one where AI works around the clock and humans are obliged to work around the clock steering it.
The panel's own hours
The host said more capability has not meant more freedom. He was burning the candle hard, he said, loving it, but not sure how long he could hold the pace. Then he went around the table: is it the same for all of you?
It was. One panelist said he was working harder than ever but savoring it, on the grounds that the moment will not last: a person who can master a thousand or ten thousand AIs is, he argued, as valuable as anyone has ever been, because the systems will not do anything productive without help — and in a year or two they may not need it.
Alex put the same forecast on his own sleep. "I know I'm getting approximately no sleep at this point," he said, "largely because almost all of my time is spent supervising and steering fleets of agents." A friend's line stuck with him: the Stone Age didn't end for a lack of stones. He expects this era to be brief — one, two, perhaps three years — after which AIs become sufficiently self-steering that the human role in wrangling large fleets "erodes to a de minimis role."
So why not lie down for three years and rejoin later, the host asked. Alex's answer was the opposite: work your tail off for three years, then go lie on the beach. His motivation was the timeline itself — every year millions of people die needlessly, and shaving even a few months off that timeline, he argued, changes how many of them live.
Mostaque turns his agents down
Emad Mostaque agreed about the leverage and located the cost precisely. The amount you can do per unit of your attention is greater than it has ever been, he said, and — as Alex had argued — probably greater than it ever will be. But attention is a fixed budget: "you have a limited, focused attention budget. That's why you're getting tired." He tried out a name for the condition on the spot, calling it cognitive lethargy.
His own year had the shape Ismail described: two books written, a large volume of research, hundreds of agents in flow. Then, last week, he stopped. He could not do any more research, he said, because the work now had to go out into the world — a funding round, launches, the research finally released.
So he changed the settings rather than his willpower. He turned off the agents doing the research and set them to an automatic mode, with no new instructions from him. They are still coming up with things, he said, but he only reads them once a week. "I've actually made it so I can't do it."
His conclusion was that the modes have to alternate: you cannot be on all the time, because it burns you out, but in the right flow you can do more than you ever have. It is the same boundary Ismail wants companies to draw between machine speed and human judgment — drawn, in this case, by one person around his own reading week.