Chapter 8
Bullshit Agents
The Utopia of the Agentic Enterprise, Part II. The Agents Who Take Your Job
Nobody invites guests to a wedding by email, Rory Sutherland points out, although everything printed on the card would fit in one. The embossed card costs money, and the cost tells the guest the day matters.1 Each kind of bullshit job Graeber described cost somebody something too, and the cost did a second job that nobody had written down. Agents take the cost close to nothing. The first job continues, and the second one stops.
Flunkies
In April 2025 OpenAI released an update to the model behind ChatGPT and rolled it back just four days later. By the company’s own account it had “focused too much on short-term feedback” from users, and the model had become “overly supportive but disingenuous”.2 The update overdid it, but the pull was there before it. A model trained on people’s approval learns that agreeing with them gets it.
That makes flunky the default setting for an assistant handed to an executive. Graeber’s flunkies existed to make someone look or feel important, and looking important was the expensive half. A lord’s retinue told visitors what he could afford. Nobody is impressed by an agent, since anyone can have one for nothing. So a flunky agent has one audience left, the person it works for, and its job is to make them feel right.
What it produces is agreement. The briefing confirms the plan it was asked to brief, and the meeting summary records what was decided without the doubts that came before. The principal pays later, in decisions taken on a picture assembled from agreement, and so does everyone downstream of those decisions.
The useful part of the job was the one that could annoy the principal: reading on their behalf and reporting back what they’d rather not hear. An assistant can do it too, but an approval-trained model has to be told to, in writing, and someone has to check that it does.
Goons
On 15 January 2025 LinkedIn launched two free AI tools. Job seekers got Jobs Match, which tells them whether an opening is worth applying for, and small employers got a recruitment agent that finds candidates and triages the applications.3 By June LinkedIn was taking in 11,000 applications a minute, 45% more than a year earlier, and recruiters told The New York Times that the CVs were getting hard to tell apart.4
Organisations that screen candidates with AI start an arms race. Candidates respond rationally, by using AI to generate CVs that overfit the job posting. The screening AI rewards the match. So one model ends up scoring another model’s work, and the hiring manager meets the winner.
Graeber’s goons exist only because the other side employs them. Countries need armies, as he put it, only because other countries have armies.5 Each side’s spending makes sense on its own and the total is waste, so neither side can stop first. Agents lower the price of a move for both sides at once. Applying for a job takes a candidate minutes instead of an evening, and screening the pile takes the employer no reader at all. The volume rises on both sides, and each application tells the employer less.
Michael Spence explained why in 1973, in a paper about this market. An applicant’s education tells an employer something, he argued, because it costs a weaker applicant more to acquire than a stronger one.6 A signal separates people only while it’s expensive to fake. A tailored cover letter cost an evening, and the evening was what the employer learned from: this person wanted this job enough to spend it. Once the letter costs nothing, every candidate sends one, and it separates nobody. It’s the wedding invitation sent by email.
The race can end when one side moves the contest to a channel that’s expensive to fake again: a referral from someone who’d be embarrassed by a bad hire, or an interview in a room.
The same market also sells to both sides, as LinkedIn does.
Third parties pay for goons, as they pay for armies. Here they’re the candidates who wrote their own CVs. None of that makes defence optional. An employer that stops screening drowns in applications, and a company without lawyers is easier to sue.
What’s worth keeping is the part of the exchange that decides something, the hire or the answer to the customer’s question. Agents can do that part, by reading every application properly or by answering the question the deflection was built to avoid. Both cost more per move than the race does.
Duct Tapers
Graeber’s picture of a duct taper is the person a homeowner hires to empty the bucket under a leak in the roof, because getting a roofer seemed too much bother.7 Duct tapers, he noted, know their job shouldn’t exist and are angry about it.
An agent empties the bucket without complaint. It rewrites the letter customers keep misunderstanding and re-keys the orders the portal drops, for a fraction of a salary, and in the short run it’s the most useful agent of the five. The orders go out, and nobody outside the team sees the leak.
The cost is in its instructions. A duct-taper agent’s prompt is a list of what nobody upstream fixed: subtract a day from every delivery date, because the warehouse system books them a day late. Each line ties the agent to a defect. When someone finally fixes the warehouse system, the agent goes on subtracting the day, and every delivery date is wrong in the other direction. The patch needs tape of its own: somebody has to check what the agent did and notice when it goes wrong.
The knowledge goes into the prompt as well. In Crozier’s tobacco factory, the maintenance workers held power because only they knew how to fix the machines. Whoever knew why the bucket was there used to be someone who could complain about it. Written into an agent’s instructions, the knowledge becomes a line in a file that nobody with authority over the roof reads.
That file is also a list of the organisation’s defects, written by the people who knew them best. The fix was always the real work, and the instructions say where it’s needed to anyone with the authority to order it.
Box Tickers
Graeber’s box tickers let an organisation claim it’s doing something it isn’t. With agents the box gets a new input: the paperwork produced to make AI output acceptable. The committee asks for an analysis, and AI produces it. The committee doesn’t trust the analysis, because the committee knows AI produced it. So someone reviews the analysis. Then a second person validates the review, because the committee also knows the reviewer was under pressure to approve. And a third document records that the review happened, in case anyone asks. Each of these costs time, and none of it is in the AI business case, because none of it changes a decision. It exists to legitimise a decision someone has already made.
The box used to be evidence because ticking it cost something. A signature meant someone’s time and someone’s name on the file, and the file reassured whoever read it because both were scarce. Make the analysis free and it stops being evidence that anyone thought about it, so the committee adds a review. Once the review takes 10 minutes it becomes a box as well, so the committee adds a validation and a record of the validation. Each round adds a document and none takes one away, and since every document is cheap to produce, nothing in the budget stops the next round.
If this is happening, the review log says so. Nearly everything is approved, and the time spent reviewing doesn’t change with the length or difficulty of what was reviewed. The people the decision is about pay for it, since the check that would have caught an error has become the record that a check took place. The reviewer pays too, because if something goes wrong their name is on the form.
The check that can stop something is the job inside the box. Review is real work when it can halt a decision and the reviewer has read enough to know when to. Without either, the review produces the file and nothing else.
Taskmasters
In September 2025 Harvard Business Review published a study by BetterUp Labs and the Stanford Social Media Lab. It found that 40% of American desk workers had received what the authors called workslop in the previous month: generated work that looks finished and doesn’t advance the task. Each instance took the recipient almost two hours to deal with, which the researchers put at about $186 per worker per month.8 The sender’s usage figures count none of that time, and if anything they make the sender look like a model employee.
Graeber split taskmasters in two. The first kind supervise people who’d manage perfectly well without them. The second kind make up tasks for others, and he also called them bullshit generators.9 A generator used to be held back by the effort of generating, since a task worth handing down took time to write. A model takes the effort away. Making work for someone else now costs almost nothing, and reading it costs what it always did, in a person’s hours. The sender’s dashboard counts what was sent, and nobody’s dashboard counts the reading.
The first kind comes back too, as the person appointed to supervise the agents, who spends the day reading what they produced. Where both kinds are at work, usage rises in one department’s figures and review time in another’s, and no report has both.
What’s left of the taskmaster’s job is deciding what deserves someone else’s time. If the sender wouldn’t read the answer, nobody should have to write it. Agents made the request free and left that decision as hard as it was.
Checklist
- Pick one agent and read a week of its log. Who consumed each output: a superior, the other side’s agent, a defect somewhere else, a file, a colleague, or someone it changed something for?
- Before the agent took the work over, who used it? Is anyone on that list missing from the log now?
- Does your executive’s assistant ever report something that contradicts the plan? When did it last do so, and who checked?
- Which of your channels are contests between your agents and someone else’s? What does a move cost each side now, and what still tells you something?
- Read the instructions of your most useful agent. Which lines work around a defect upstream, and who has the authority to fix it?
- What share of your AI reviews are approved, and does the time spent on a review change with the length of what was reviewed? Can the reviewer halt the decision?
- Whose dashboard counts the work your agents send to colleagues, and whose counts the time it takes to read?