I Run a Publishing Company Where AI Agents Do Almost Everything. Here Is the Job They Cannot Touch.

I run an independent publishing company. It puts out nonfiction and fiction: AI and technology guides, ancient history, cybersecurity, productivity, fiction. It has a real catalog, real revenue, and a real production pipeline. And nearly all of the production work is done by AI agents.

This is not a demo I built to prove a point. It is an operation I run week in and week out, and it sits on a boundary that anyone building with agents is going to hit. Everything below a certain line, my agents handle so well it has become boring. Everything above that line routes through me. Every single time. Figuring out why the line sits exactly where it does turned into the most useful thing I have learned about agentic systems, so let me show you both sides of it.

What the agents do

Watch how little of this involves me.

A new book starts as a one-line idea. An agent develops it into a full proposal: premise, audience, chapter outline, positioning against comparable titles, marketing angles. The proposals are good. Not "good for a machine," just good, the kind I would have been happy to write myself, produced in minutes instead of days.

Then the pipeline runs. An agent expands the outline into a table of contents. The chapters get written, and here the system does something no human author can: it writes them in parallel. Several agents each take a batch of chapters, write them in the book's voice against the outline, and save them as they go. A full manuscript that would take me months materializes in hours. The chapters cohere, hold the style guide, and respect the arc of the book.

Then assembly. The chapters get stitched into a single manuscript with front matter and navigation, exported into the EPUB and DOCX formats the retailers want, correctly formatted for Kindle. Cover art gets generated in a few variations with the title and author composited on. Then the marketing layer, which is its own mountain in traditional publishing: retail description tuned to the platform's rules, categories and keywords, launch checklist, social posts, listing data laid out so the retail forms are nearly copy-paste.

The whole way through, the pipeline tracks itself. Every book has a state record in a database, kept in sync with the files on disk. If a run gets interrupted, it resumes where it left off. A book goes from idea to publish-ready, manuscript through marketing package, in single-digit hours, most of it unattended. Five years ago that was a team and a quarter of a year.

I am telling you this in detail for one reason: so you cannot dismiss what comes next as the complaint of someone whose agents are not good enough. My agents are extraordinary. The task layer, in my business, is essentially solved. Which makes the shape of what is left over very easy to see.

What the agents never do

Here is the work my agents have never done. Not once, not partially. Never attempted, because nothing in the system points at it.

They never decide which book to write next. The pipeline is world-class at making the book I chose. It has no opinion about what I should choose. The single most consequential decision in a publishing company, what to publish, has never been touched by the machinery that does everything else.

They never notice a genre is saturated. If a category fills up with competing titles and the economics quietly die, no agent tells me. The pipeline will keep producing beautiful books into a dead market with total competence and total indifference.

They never reallocate effort. When one category fades and another grows, moving effort from the fading line to the growing one is a decision I make and hand down as new tasks. No agent has ever seen effort as a spendable, steerable quantity.

They never kill a series. Deciding that book four should not exist, no matter how well the pipeline could produce it, is pure judgment: sunk costs, opportunity costs, a loss to take. Agents will cheerfully generate the proposal for book four.

They never propose changing a price. Pricing is numerically trivial and strategically loaded. No agent in my system has ever suggested touching it.

Notice what that second list has in common besides its contents: none of it is urgent. A chapter that needs writing announces itself. A genre that is quietly dying announces nothing. Tasks fail loudly. Responsibilities fail silently, until months later you realize the catalog drifted somewhere you never chose to go.

It is not that the models cannot reason about this

Here is where almost everyone guesses wrong. The obvious answer is that those decisions are too hard for AI, that reading markets and setting prices needs judgment the models do not have.

That is not the blocker. If I paste my sales data into a frontier model and ask what it would do, I get genuinely competent strategic analysis: plausible reads on which categories look tired, sensible pricing arguments, reasonable proposals for what to publish next. Is it right? Sometimes. Strategy is probabilistic, and so am I. But it is at least as defensible as my own reasoning, and mine is what currently makes every one of these calls. The raw cognitive work of publishing strategy is within reach of models that exist today. You can verify that yourself in an afternoon.

So why is none of it delegated? Because nothing in the system is built to be trusted with those decisions. Look at what is actually missing, because it is not intelligence, it is plumbing.

There is no verified feedback loop the system can act on. Sales live on retailer dashboards built for human eyeballs. My agents can write a book, but they cannot watch it sell. There is no trustworthy, machine-readable stream connecting market outcomes back to the system that produced them. I am, functionally, the company's sensory organ, carrying the signal from the dashboards to the pipeline by hand.

There is no budget authority. My agents have never spent a dollar on their own judgment. There is no wallet an agent controls, no mechanism by which the system could decide "this niche deserves five more titles" and fund it. Without resource authority, there is no resource allocation, and without allocation, no ownership.

There is no reversibility for a bad strategic call. This is the asymmetry that really governs the line. When an agent writes a bad chapter, the blast radius is nothing: I regenerate it, minutes and pennies, so I delegate chapter-writing without a second thought. But a mispriced catalog, a dozen titles dumped into a dying genre, a series killed right before it turned, those cost months and real money, and some cannot be rolled back at all. I have cheap undo at the task layer and no undo at the strategy layer. So the strategy layer stays human. The principle hiding in that sentence is the one I keep coming back to: delegation follows reversibility. You hand over exactly as much as you can afford to have done wrong.

There is no audit trail for judgment. When my pipeline produces a book, I can reconstruct every step. If a strategic layer existed, could it show its work the same way: why it chose this niche, what it weighed, what it expected versus what happened? Nothing like that exists. Not because it is impossible, it is a database schema and some discipline, but because nobody has built it. And without it, "trust the system's strategy" means "trust, unverifiably."

Four missing pieces: verified feedback, budget authority, reversibility, auditability. Not one of them is a model capability. They are infrastructure. The boring stuff. The gap between my agents and my company is not a gap in anyone's benchmark scores. It is a gap in accountability plumbing that has simply never been built.

The bottleneck moved into my head

So here is the honest org chart. At the bottom, a production layer made of agents: fast, parallel, tireless, improving every quarter. At the top, a responsibility layer made of one human. I own the objectives, sense the market, rank the tasks, hold the budget, absorb the failures, and revise the strategy.

The bottleneck in this company used to be production. Writing was the slow, expensive part, so tooling ate it. Now production is nearly free and the bottleneck has moved up to exactly where I stand: the deciding. The production layer scales horizontally; if I wanted five books in flight instead of one, the pipeline would not blink. The responsibility layer does not scale at all, because it is one person's attention. When people ask why my catalog is not ten times larger given what the pipeline can do, this is the honest answer. The constraint was never how fast books can be made. It is how fast one responsibility layer can decide, verify, and answer for what should be made.

The industry keeps asking when the models will be smart enough to run a business. Standing inside a business that agents already mostly run, I can tell you it is the wrong question. The models can already reason about my strategy about as well as they write my chapters. What they cannot do is be trusted with it: verifiably, boundedly, reversibly, auditably. Trust, not intelligence, is the bottleneck. That claim is the hinge of the whole argument.

The book

This is the case study at the center of my new book, The Responsibility Layer. The chapter you just read the bones of is the evidence; the rest of the book is the argument it supports, and where it goes next: what the "responsibility layer" actually is, why trust rather than intelligence is the real constraint on autonomy, and how to start building the accountability stack that would let you delegate judgment the way we already delegate tasks, small and reversible first, authority growing with track record.

If you are building with agents and you have felt this same line in your own work, the place where the automation stops and you start, the book is for you. You can find The Responsibility Layer on Amazon here: The Responsibility Layer - By Shane Larson

It comes out of Peak Grizzly Publishing, the company described above, which is itself the experiment. There is something fitting about a book on agentic production being produced by an agentic pipeline. The one part no agent wrote is the part the book is about.

Disclosure: the Amazon link above is an affiliate link.