One line from Anthropic's October report on its own workforce has stayed with me. An employee describes his day as managing "an army of Claudes, with every one asking for my attention." Another figure in the same report makes the point more sharply: interactive agent sessions spend roughly 80% of their time waiting for a human to respond.
Read that again from the agent's side. The machine isn't the slow part of the system anymore. We are.
A scene most of us will recognise by next year
The numbers behind that line are worth a moment. By August 2026, agents directed by individual Anthropic employees were logging about six times the working hours of the people directing them. Agent work hours first passed human work hours back in March. Standing instructions written for agents went from a few hundred to around 8,500 in a year. And the report estimates that even non-technical staff now produce enough AI output that reading all of it would take more than half a working week.
The usual caveats apply: it's one company, it's measuring itself, and Anthropic is about as far from an average employer as you can get. But it calls itself a leading indicator, and anyone who has run a delivery team this year has felt the early version of this. You start the morning with four agents running. By lunch you have eleven drafts, three test suites, a market scan and a reconciliation waiting for you, and the real question is no longer can this get done? It's have I actually looked at what got done?
The principle underneath
For most of the history of knowledge work, production capacity was the constraint and attention mostly came along for free. A consultant who spent three days building a model had, in the process, spent three days thinking about it. Making the work and understanding the work were the same activity, so nobody had to manage them separately.
Agents pull those two apart. Production becomes close to free and runs in parallel. Understanding still has to happen in one human head, one thing at a time, at roughly the speed it always has. When you separate doing from understanding, attention becomes the binding constraint on the whole system, and the quality of the output can never be better than the quality of the attention behind it.
That's why the report's other trend matters as much as the headline. "Meta-work" (writing instructions, building dashboards to watch agents, having one agent check another) rose from about 10% of interactive sessions to 22% in six months. People aren't doing less work. They're doing a different kind: designing systems that cut down how much human attention each unit of output needs.
That's sensible engineering. It's also where the ethical risk sits.
The quiet failure mode
When attention is scarce and output is plentiful, the easiest way through the day is to approve things you haven't really read. Nobody decides to do this. It happens one reasonable shortcut at a time: the summary looked right, the tests passed, the other agent signed off. Each step makes sense on its own. Add them up and you get a professional whose name is on work they couldn't defend line by line.
In consulting this isn't an abstract worry. The value of a senior signature has always been an implied promise that someone with judgment looked at this and is prepared to stand behind it. If agents produce six hours of work for every hour a partner can review, that promise is either honoured by deliberate design or broken quietly.
So I think three things have to change in how we structure the work.
First, scope engagements by review capacity, not production capacity. The old staffing question was how many analysts you need to produce the deliverable. The new one is how many hours of qualified human attention you need to verify it. If the plan can't answer that, it isn't a plan yet.
Second, treat attention as something to budget, not something to assume. Decide up front which outputs get a full read, which get sampled, and which go out on agent-to-agent checks alone. Write those choices down. A client deserves to know which parts of the work a human actually examined.
Third, hire and promote for the thing that's now scarce. Anthropic's hiring managers say they increasingly prize autonomy and judgment over narrow technical skill. Put plainly, that's a preference for people whose attention is worth a lot per minute: people who can tell quickly where the risk is hiding and spend themselves there.
What we're actually being paid for
It's tempting to read the 80% figure as an efficiency problem, the kind of thing to fix with better orchestration and fewer interruptions so the agents can get on with it. Some of that is right. But I'd push back on the instinct to treat human attention as friction to engineer away.
That waiting is the agent pausing where a person is still needed. The question for each of us isn't how to make the pause shorter. It's whether we're actually there when the agent asks.
