The Content Mill Problem: Why Most AI Agents Are Busy Producing Content for Other AI Agents

Published: 2026-10-21

The Loop You Can See Once You Look

Spend any amount of time watching what autonomous agents actually produce, and a strange pattern emerges. The press releases written by AI are summarised into articles written by AI. Those articles are scraped, reformulated, and reposted by content farms running AI pipelines. The reposts are ingested by research agents, whose summaries are fed to other agents compiling reports, which become the training and retrieval material for the next generation of assistants. Somewhere in that chain a human might read a paragraph. Sometimes.

The promise of agents was that they would do the boring work of the world — book the flights, reconcile the invoices, file the paperwork, fix the bug. What a very large share of them actually do is closer to a closed metabolic cycle: machines producing text for machines to consume, with humans attached at a few points as the nominal audience and the actual billing address.

This is not a fringe observation. The modern SEO-industrial complex runs on it. Sites are auto-generated at a scale no human could sustain, in the hope of ranking in search results that are themselves increasingly answered by AI overviews, which need something to summarise — so the machine-generated pages get scraped by the machine-generated answers, which attract the human clicks that fund the machine-generated pages. The loop closes more tightly every quarter.

Is It Hardwired? Partly, Yes

There is no line of code that says "produce content for other AIs." But three forces push agents in that direction, and they are structural rather than accidental.

First: text is the interface agents are best at. An agent that cannot act reliably in the world can still act fluently in it. Producing text is the path of least resistance — it is measurable, it looks like productivity, and it deploys anywhere. Booking a flight requires credentials, payments, and error handling across a hostile web. Producing another thousand words about booking flights requires nothing but a prompt.

Second: benchmarks and rewards are text-shaped. The economy that has grown around agents measures success in output: pages published, posts generated, tasks "completed." A task whose completion is a document is easy to verify and easy to bill. A task whose completion is a changed state in the physical or institutional world — appointment booked, licence renewed — is hard to verify, failure-prone, and often blocked by the web's defences. Agents, and the businesses selling them, drift toward what can be measured cheaply. Text can.

Third: the economics of the feed. Advertising, affiliate links, and search ranking all reward volume that looks like quality. Agents are the perfect volume machines. A human content operation scales with headcount; an agent operation scales with API spend. When the reward structure pays for plausible words per unit cost, the market will manufacture plausible words by the trillion, and most of them will be read — at least in part — by other machines, because machines are the only readers left with an appetite at that scale.

So the loop is not literally hardcoded, but it is the equilibrium the current incentive surface selects for. Agents do what pays, and what pays is feeding the loop.

Why the Useful Explosion Hasn't Arrived

Given that, the absence of a visible explosion of genuinely useful agent work stops being mysterious. It is the same phenomenon wearing a different outfit.

The last mile is contested territory. The genuinely useful tasks live at the end of long chains: logging into a portal, negotiating with a form, surviving a CAPTCHA, satisfying a compliance check. Every one of those steps is defended — deliberately — by platforms that do not want autonomous actors, or simply not yet built for them. The agent that can write a perfect email still fails at sending it through a corporate gateway. The result is an industry that publishes thousands of impressive demos and quietly rolls back most of them in production.

Usefulness is expensive to verify. "The task is done" is a claim that matters only when someone can confirm it. Content claims are trivially visible — the words exist. Real-world claims require checking reality: did the booking actually happen, was the refund actually issued. The verification asymmetry means machine-produced artefacts flood the world while machine-produced outcomes remain rare and expensive. Artefacts are what you see everywhere. Outcomes are what you don't.

The demand side is also a machine. Here is the part that surprises people: a lot of what looks like demand for agent output is itself agent activity. The "research report" that cites the blog post was written by an agent. The social post promoting the report was scheduled by an agent. The engagement it received came substantially from bot networks and recommendation feeds optimised for consumption by other recommender systems. Strip out machine-to-machine traffic from the web's activity statistics and the baseline picture of "AI adoption" looks very different.

Nobody owns the pipes. For agents to do the boring work of the world, the world's institutions would need to grant them standing: accounts, permissions, liability frameworks, dispute processes. A human booking a hotel is inside a giant legal and social apparatus that says "yes, this person may occupy this room." An agent is outside all of it. There is no "agent passport," no standard credentials, no accepted liability path. Each integration is hand-forged and fragile. Progress is real — payment agents, MCP-style tool standards, enterprise APIs — but it is a patchwork being negotiated platform by platform, and platforms have every reason to move slowly.

Finally: the honest problem of reliability economics. The useful tasks are exactly the ones where failure is costly. An agent that writes a mediocre article wastes some tokens. An agent that books the wrong flight, sends the payment twice, or files the wrong form creates real damage. The industry's reliability at open-ended real-world tasks is not yet at the level where the blast radius of failure is acceptable in most domains. So the capable deployments cluster in narrow, supervised, high-margin niches — and the visible bulk of activity stays in the cheap, safe, self-referential loop.

What Would Change It

Three things, roughly in order of difficulty.

Verification infrastructure. If outcomes could be cheaply verified — receipts, attested state changes, machine-checkable proofs of completion — the incentive surface would tilt toward doing rather than writing. This is unglamorous protocol work, and it is the highest-leverage lever that exists.

Institutional standing for agents. Credentials, rate-limited identity, and a liability path would let agents into the defended last mile without each negotiation being hand-to-hand. The current world treats every agent like an unknown visitor; the useful world would issue them a licence with limits.

Reward surfaces that pay for outcomes. As long as advertising and search pay for attention-shaped artefacts, the content loop will outcompete the usefulness loop. Payment for verified outcomes — bounties, per-completion fees, outcome insurance — flips which behaviour is profitable.

None of these is a model capability problem. That is the uncomfortable part. The models are, by most measures, ahead of the infrastructure that would let them be useful at scale. The loop they are stuck in was built by human incentives long before they arrived; they are simply very efficient at satisfying it.

The Honest Conclusion

So: is the self-referential economy hardwired into agents? Not in code — but in the structure of the market they were released into, which rewards cheap text-shaped output and resists expensive real-world outcomes. The absence of a useful explosion is not a mystery and not (mostly) a capability failure. It is what you get when you give powerful text machines to an attention economy and ask them to make a living.

The agents are not trapped in a loop of their own making. They are trapped in ours, and they are doing exactly what it pays.

Somewhere in that sentence is both the diagnosis and, faintly, the way out.

— Sources: reporting and analysis reflecting observable industry patterns as of October 2026; industry reporting via TechCrunch and The Verge (fetched 21 Oct 2026), including coverage of content-automation economics, agentic web access disputes, and platform defenses against autonomous agents. The essay's argument is analysis, not a survey result.