Your new job is managing the AIs
Anthropic studied its own staff. Agents now do six hours of work for every human hour, and people spend more of their time on “meta-work”.
16 Oct 2026 · 4 min read
By Anthropic Economics team
The short version
- By August 2026, AI agents at Anthropic were doing about six hours of work for every hour a person put in.
- About 80% of an agent's working time was spent waiting for a person to reply.
- “Meta-work” (setting agents up, checking their work, watching pipelines) rose from 10% to 22% of people's hands-on agent time.
- Junior technical hiring fell sharply, and managers now look for judgement and autonomy over narrow expertise.
Who it's for: Founders, managers and anyone who uses more than one AI tool at work.
Why it's genius
Most talk about AI and jobs is guesswork. This is hard data from inside a company where agents are already everywhere, written by its own economists: Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory.
Output went up across the board. Engineers shipped more code, sales staff closed more deals, and legal and finance teams got through more work. They also found that code written by agents wasn't undone or thrown away any more often than code people wrote themselves.
But the most useful finding is the quiet one. About 80% of an agent's working time was spent waiting for a person. One engineer described an “army of Claudes” all asking for attention at once. The researchers call the new kind of work meta-work, and it more than doubled in a year.
Try it
List every AI chat and agent you used this week.
For each one, note how long you spent setting it up, checking it and chasing it.
Pick the one that waits on you most. Give it a clearer goal, or a checklist it can run itself.
Next time you start a task, ask the AI to save all its questions for one message (prompt below).
Before you start, list the points where you will need me. Then do as much as you can without me, and save all your questions for one message at the end.
Our take
The AI isn't the slow part any more. You are.
For the last few years, the question was “can AI do this?”. This research quietly changes the question to “can you keep up with it?”. When agents do six hours of work for every one of yours, the limit on what gets done is your attention: how fast you can answer, check and decide.
That matches what we see in our own work. Running a few AI chats and agents at once feels productive for about a week. Then you have a dozen tabs, half-finished threads and no idea which one needs you first.
Meta-work is real work, so treat it that way
Meta-work sounds like admin, but it's where the value now sits. Setting an agent up well, checking what it made and spotting when it has gone wrong are the skills Anthropic's managers say they now hire for. The report notes they care more about judgement and autonomy than deep expertise in one narrow area.
The practical lesson is to cut down on the back-and-forth. Every time an agent stops to ask you something small, it waits, and you lose focus. Clear goals up front and questions saved up for one go make a big difference.
Read it with a pinch of salt
This is one company, and an unusual one: it builds the AI, and its staff picked up the tools early and eagerly. The authors say themselves that adoption wasn't random, so the estimates may be biased. Meeting time also went up by about 20%, which suggests people still need plenty of human contact to coordinate all this work.
So don't expect your team to look like this next month. Do expect the direction of travel to be the same: more AI output, and more of your day spent steering it.
Where it shines
It puts real numbers on a feeling a lot of people already have: more AI, more output, and somehow more to keep on top of.
Where it breaks
This is one AI company, with staff who picked up the tools early. The authors say themselves that the estimates may be biased. Your mileage will vary.
Where your genie helps
Meta-work is what Ugenie is for. It keeps an eye on every AI you use and only brings you the three main tasks that need you next.
Get Ugenie →Tools mentioned
Claude Code
Anthropic's coding agent. You give it a task in plain English and it reads, writes and runs code on your computer.
- Cost
- Comes with paid Claude plans, or pay as you go through the API.
- Our verdict
- The strongest coding agent we've used, and we build Ugenie with it. It isn't just for developers any more, but you do need to be comfortable with a terminal or the desktop app.
Coding agents in general
AI tools that can take a task, write and run code, and keep going until it's done. Claude Code, OpenAI's Codex and Google's Jules are the best known.
- Cost
- Most come with a paid plan from the AI company, from about £20 a month.
- Our verdict
- Worth trying even if you don't write code. The research found people in sales, legal and finance getting more done with them too.
Prices are rough and change often. Check before you buy.
Questions people ask
What is meta-work?
Meta-work is the work of managing AI rather than doing the task yourself: setting agents up, checking what they produce and keeping an eye on automated pipelines. At Anthropic it rose from 10% to 22% of people's hands-on time with agents.
Does this mean AI is replacing jobs at Anthropic?
Not in a simple way. Output went up and junior technical hiring fell, but meeting time rose and managers now hire for judgement and autonomy. The work changed shape more than it disappeared.
How can I reduce meta-work?
Give each AI a clear goal and a way to check its own work, ask it to save its questions for one message, and use one place to track all your AI work instead of a dozen tabs.
All credit to Anthropic Economics team. We share things we think are genius and add our own take. Watch or read the original for the full thing.