Upload
Oct 19, 2026
7 Views
0 Comments
0 0

Human attention is the new old bottleneck. We have to design for it even more.

Written by
A wide office floor where one person reads at a desk while dozens of screens around them keep filling with new work.
A wide office floor where one person reads at a desk while dozens of screens around them keep filling with new work for them to address.

Agents at Anthropic now log six hours for every human hour, and reading their output would eat half a week. The job is to create judgement systems for what deserves attention.

On October 16, four researchers at Anthropic published How employment is changing at Anthropic, a study of how its own agents changed the work.

Anthropic is one of the companies at the front of this shift, running its own work on the models it builds, so the study is an early read on where other companies are heading.

And honestly, it looks like previous studies in new clothing as I have linked below, just accelerated. Read on.

Agents work more than humans. Not a surprise.

Here are the early results: By August 2026, agents directed by employees were logging six hours for every hour their people worked.

Agent sessions spent about 80% of their time idle, waiting for a person to answer. Even for employees who don’t write code, reading everything their agents produced would take more than half the workweek.

That’s the reading tax, and every team with agents will pay it.

None of this is new — we have decades of research and many, many principles like Automation Bias that we can apply to solve these problems. Now it’s up to us to apply them.

This productivity is changing for everyone.
This productivity is changing for everyone.

Underneath the numbers, the nature of work is changing yet again, and this time it’s turning everyone into a manager instead of a maker of widgets. That makes this a story about human in the loop and who keeps the agency. A person who can’t read what an agent did is still in the loop on paper, while the agent makes the calls.

That puts the problem on ground designers, product managers, and researchers already work.

Here are seven steps for keeping people in charge of that loop.

Measure Reading Time

Herbert Simon saw this coming in 1971. In Designing Organizations for an Information-Rich World, he wrote that “a wealth of information creates a poverty of attention,” with most of the cost landing on whoever receives it. He meant data banks, and the line fits agents better.

Gloria Mark at UC Irvine has measured how thin it already is.

Her studies from 2016 to 2020 found people stayed on one screen for 47 seconds on average. Her first study, published in 2004, found two and a half minutes, according to the University of California’s How to sharpen your attention and meet your goals in 2024.

Every agent you add is a new sender in an inbox that was already overflowing with more information.

Note: Hours estimated by converting output tokens to words, assuming 0.75 words per token and 250 words read per minute. What’s lost in here is that a lot of work goes unreviewed, no human in the loop.

Anthropic measured the agent side the way a designer would, converting each session’s output into words read at 250 words a minute.

Even non-technical groups crossed half a workweek of reading.

Most days I build now, with a coding agent doing much of the work, and my limit is how much I can read and judge. So treat reading time like page load time, a budget a team sets and defends.

What’s lost in here is that a lot of work goes unreviewed, no human in the loop. Four hundred and twenty four hours is a lot more than the 168 a week currently has.

If one workflow’s weekly agent output, divided by 250, comes to more hours than anyone has, the workflow has a design problem.

Rank Agent Requests

One Anthropic engineer compared managing agents to babysitting. A child who needs a juice box opened and a child with a cut finger both arrive at full volume, and so do the agents.

Hospitals learned what that costs.

Joint Commission’s 2013 Sentinel Event Alert on medical device alarm safety estimated that 85% to 99% of alarm signals don’t require clinical intervention. Staff stop hearing them which is alarm fatigue in spades. Its database logged 80 alarm-related deaths between January 2009 and June 2012.

An agent that flags everything at the same volume has trained its human to ignore it.

Hospitals answered with alarm management: priorities, quiet defaults for routine readings, and a named person who responds. Mark Weiser and John Seely Brown at Xerox PARC described the goal in The Coming Age of Calm Technology in 1996, as technology that moves from the periphery of attention to the center and back.

Most agent interfaces skip the periphery and drop every request in the center.

Make agents justify a severity level for each request, keep the routine ones in the periphery as a digest, and bring only decisions a person must make to the center.

At Anthropic, agents mostly post in channels set aside for Claude-human work, which keeps the threads where people talk readable.

Design the Review Step

Review is where people pay the reading tax, and it’s where a human in the loop either keeps agency or hands it over. At Anthropic, agent-written code was removed or reverted at the same rates as code the same engineers wrote by hand, though the authors caution that agent errors may be harder to spot.

Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity

Perception is the bigger risk. In Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, a 2025 trial, METR found that 16 experienced developers took 19% longer with artificial intelligence tools after expecting to save 24%.

If people can’t tell how long review takes, they can’t tell when they’ve stopped doing it.

Andrej Karpathy made the same point in Andrej Karpathy: Software Is Changing (Again), his June 2025 keynote for Y Combinator. Models generate and people verify, he said, so the loop between them should run as fast as possible, with the model kept “on the leash.”

My Automation Bias entry on UX Guidelines covers what happens when that cycle slackens: people approve what the machine proposes and miss what it never flagged.

A review screen should show what changed next to what existed, what the agent checked and skipped, and where each claim came from.

Put the riskiest change first, and log review time per item so the team can see the tax.

Write Decisions Down

A manager’s decisions only count if the agents can see them. Agents reason from whatever people leave in the open. One Anthropic employee’s name stayed on a months-old document about company swag, so agents kept sending them requests for T-shirts and mugs long after the job changed hands.

Shopify’s chief executive, Tobias Lütke, gave this work its popular name in June 2025. He preferred “context engineering” to prompt engineering and called it the core skill, as Simon Willison recorded in Context engineering.

Anthropic’s engineers add, in Effective context engineering for AI agents, that a model’s context is a finite attention budget. Every loop runs on two budgets, the agent’s context and the person’s reading time.

The cheapest way to protect a person’s attention is to write the decision down once, where every agent can find it.

At Anthropic, written standing instructions for agents grew from a few hundred to 8,500 in a year. Meeting time grew 20% for the median employee and agents see none of a meeting unless someone writes it up.

For design and research teams, this is familiar work. Decision logs, research repositories, and design rationale used to be documentation few people opened.

Now agents read every page, which raises the bar for keeping it current and owned.

Manage the Agents

At Anthropic, a fifth of employees surveyed in April 2026 said their role had shifted from doing the work to managing Claudes, and interviewees described individual contributors making VP-level decisions. Peter Drucker saw the first version of this in The Effective Executive in 1967, writing that “every knowledge worker in a modern organization is an ‘executive’” if they’re responsible for a contribution.

Agents extend that to anyone with a login, and Microsoft’s The 2025 Annual Work Trend Index: The Frontier Firm is born has a name for the role: agent boss.

I call it the work about the work, and the report calls the supporting work meta-work: specifications for multi-agent runs, loops where one agent reviews another, and pipeline dashboards.

It grew from 10% of employees’ interactive sessions in fall 2025 to 22% in spring 2026. Ethan Mollick argues in Management as AI superpower that this is ordinary management. What’s scarce, he writes, is “knowing what to ask for.”

Every one of those homemade dashboards is a product nobody designed.

Christopher Noessel argued for designing this layer in Designing Agentive Technology in 2017. Anthropic’s content team shows where it leads: the writers behind the greetings on Claude.ai taught agents to write that copy, then moved into writing prompts and evaluations.

Designers and researchers already have a manager’s skills here; the job is to claim this layer before it hardens into internal scripts.

Hire for Judgment

At Anthropic, an automated review of interview scorecards found that mentions of candidates’ creative use of artificial intelligence tools roughly tripled for technical roles and quadrupled for non-technical ones between 2024 and 2026. For technical candidates, comments on autonomy and judgment rose the most, while domain expertise shrank as a share of the assessment. In effect, Anthropic is hiring managers of agents.

Kent Beck, who created extreme programming, saw the split early. In April 2023 he wrote 90% of My Skills Are Now Worth $0, with a subtitle saying the other 10% were worth 1,000 times more.

A portfolio of polished screens shows what a model can now produce, and a record of what you rejected shows judgment.

The World Economic Forum’s Future of Jobs Report 2025 found that seven in 10 employers consider analytical thinking a core skill, the top of its list. For anyone in design, research, or product, that changes what goes in a portfolio and what happens in an interview.

Show the decision trail: what the agent proposed, what you cut, and why the result held up with users.

In interviews, hand candidates a page of agent output with three planted problems, and watch which ones they catch first.

Rebuild Entry-Level Roles

The hardest number in the report concerns people never hired. Junior employees’ share of Anthropic’s technical departments slipped from 21% to 14% since early 2025, and junior technical hires fell from 27% to 9%.

Anthropic isn’t alone.

This is a warning sign.
This is a warning sign.

In Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Erik Brynjolfsson and two Stanford colleagues found a 13% relative decline in employment for 22- to 25-year-olds in the most exposed occupations. Not every executive agrees.

The Register’s AWS CEO says using AI to replace junior staff is ‘Dumbest thing I’ve ever heard’ covered Matt Garman’s argument from August 2025. Juniors cost the least and use the tools the most, he said, and without them nobody will have learned anything in 10 years.

The skill to supervise an agent comes from years of doing the work it now does.

Lisanne Bainbridge warned of this in Ironies of Automation in 1983: automate most of a job, and people keep its hardest part, the monitoring, without the practice that built their judgment. Entry-level roles in design and research can be rebuilt around review, with a junior making the first call on agent output and a senior explaining every override.

That’s an apprenticeship, and it costs less than hiring a senior who never had one.

Conclusion

In 1950, the United Auto Workers and General Motors signed a five-year contract that Fortune magazine called the Treaty of Detroit. Walter Reuther’s union gave management room to run the plants, and workers got cost-of-living raises, an annual wage increase, pensions, and health benefits.

The report reaches for that same settlement and names Anthropic’s version of it on three points:

  • Broad stock ownership
  • An early warning system for jobs at risk
  • A priority on moving people into new roles inside the company

Most of us don’t set compensation policy. We do set what agents show people, when they interrupt, and how review works. Those choices decide whether everyone’s new job as a manager comes with real agency or just a sign-off button.

Right now whoever builds the dashboard is making them, and designers, researchers, and product managers should be.

The study’s framing is careful. Anthropic is one unusual company, growing fast and building the models it uses, so its practices won’t transfer cleanly to anyone else. The direction is still clear enough to act on this quarter. Output keeps getting cheaper while human attention stays fixed, and the people who decide how it gets spent will shape what these teams become.

Action items

  • Measure one workflow. Count the words its agents hand a person in a week, divide by 250, and compare the hours with what anyone has.
  • Tier agent requests. Give every request a severity the agent must justify, batch the routine ones into a daily digest, and keep interruptions for decisions only a person can make.
  • Log decisions publicly. After any meeting that changes direction, post the decision and its owner where agents and colleagues both read.
  • Pair juniors on review. Let a junior make the first call on agent output, with a senior explaining every override.

Resources

Human attention is the new old bottleneck. We have to design for it even more. was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

Article Categories:
Technology

Recent Articles

Related Articles

Leave a Comment