Slop Mode

Most people are using agents to do the same work, just faster. That's slop mode. Same output, a tenth of the cost. Hours down, ceiling flat. The opposite move: spend the saved capacity on the version you would have cut. Not "ship faster." Ship the thing that would not have been on the spec.

Two Modes

There are two ways to use an agent.

Mode one produces metrics. Mode two produces artifacts. "We ship 3x faster" is a number you can put in a deck. "Our output is better" is a claim about a thing people feel. Left to defaults, the speedup gets reported as cheaper, never as more ambitious.

The trap isn't just that the work gets cheaper. It's that the savings trigger a new mandate: half the headcount, double the volume.

And when you horizontally scale an agent that defaults to "average," the regression compounds. It isn't just a linear increase in output; it's a structural collapse in quality.

When the goal becomes more (more tickets, more emails, more roadmap), the human judgment that used to catch the corners is drowned by the sheer volume. Atlassian doesn't ship a worse Jira on purpose. It ships the same Jira, faster, with one fewer reviewer in the loop.

That's slop. Not garbage. Generic. The product no one feels when they touch it.

Mode two can avoid this, but only if the priorities are right. The goal has to be better, not more. Rethinking ambition means shipping fewer, more deliberate things. The reviewer becomes the rate limit on purpose. The volume that broke the review loop in mode one is gone. The product gets easier to test because there is less of it, and what's there is structurally superior.

I'm a product engineer by trade. Now I'm building a personalized audio story app for four-year-olds: their name in the story, their rabbit lovey along for the ride, a memory layer that remembers they helped a little star find its way home two weeks ago. So they feel a little safer in their world. Technical capability pointed at a feeling.

What Corporate Cannot Do

Inside a company, AI agent savings are predictably spent in two places: padding the margin and accelerating the committed roadmap. Ambitious bets are banished to the edge: innovation labs, demos, skunkworks. They are flashy, but they are also the first budgets to bleed out when shareholder scrutiny arrives.

The core product is a different beast entirely. It operates on entrenched revenue expectations, sales quotas, roadmap debt, and a rigid margin model. That machine knows exactly how to absorb AI as an efficiency tool: same output, fewer people, lower operating costs. What it cannot tolerate is AI as a disruptor. To use AI for radical product shifts (more bespoke care, deeper review loops, entirely new value propositions) requires cannibalizing the current cash cow. If an agent does the work of ten users, it eats ten paid SaaS seats. If it solves problems instantly, it cannibalizes the premium support tier. Corporate can fund the demo. The core business is paid to kill it before it destroys their metrics.

Kodak invented the digital camera in 1975 and shelved it. The camera would have eaten the film business, and film was the margin. A public-company CEO who voluntarily destroys the cash cow doesn't survive the next earnings call. The reimagining of photography came from a phone, because the phone-makers had no film business to protect.

Same shape with agents. The work that would have been unreasonable to staff before won't get built there. It only runs on your own.

The Floor and What We Reach For

Read both. The first is single-shot AI output. The second is the ceiling: a pipeline of validation gates with human feedback at the right moments. The difference is not polish. It's whether the child does anything.

Around the next bend was a wild strawberry — but down here, it was the size of a mountain. Red as a fire engine, dotted with tiny yellow seeds. Ayla pressed her face against it and breathed in deeply.

She broke off the very tiniest piece and ate it. It was the sweetest thing she had ever tasted. She saved the next tiny piece for Pepper, who she balanced very carefully on the strawberry leaf so Pepper could smell it properly.

Notice what the child does. She smells. She eats. The world provides; she receives.

A pebble path stopped at one empty space in the dirt. Three ants waited in a line, the first carrying a crumb taller than Ayla's knee.

Pepper looked at the ants. Then at Ayla. "They are forming a very small traffic jam."

Ayla found a loose pebble and set it in the empty place. It rocked left. It rocked right. Then it tipped out and rolled against Pepper's foot.

Pepper froze. "The hill has returned to me."

Ayla picked up the pebble again. This time, she didn't push. She listened. Her Nani and Khalas listened to her like that, when she had something hard to say. Tap-tap on one side: a flat sound. Tick on the other: brighter. She turned the pebble, dusty side down, warm side up, and tried again.

Plink.

The sound ran down the path. Plink, plink, plink.

She tries. It fails. She listens. She turns the stone. The plink is earned, not given. Her Nani and Khalas listen to her like that. The story remembered.

Twenty stories for one child, each pulling forward from the ones before. Last Tuesday's brave moment gets called back when tonight's bedtime gets hard. Over a month, the system stops sounding like a content engine and starts sounding like someone who remembers her. That is the product.

Corporate cannot do this. Per-child review loops cost per-child money. The margin model won't justify the spend. The savings are already booked.

What Doesn't Delegate

AI does not replace human judgment. It displaces it from the middle of the work to the edges. Three edges, specifically.

Taste. Someone has to know what good looks like before the machine runs. An agent with no taste-bearer upstream produces competence without distinction: output that reviews fine because nothing is technically wrong. I've read 300 agent drafts this year, and I can tell you which word shows up when the model is bored: "warm." Twenty-one times in one draft. I counted. If no one is counting, the word wins.

A tight feedback loop. Agents drop the cost of iteration close to zero. The new rate limit is how fast and precisely a human reads the last output and directs the next. The intuition that fired after long build intervals is now the interface. (I'm writing up the full working model in a follow-on post.)

Knowing when to push in. Most of the work the agents do, I do not touch. But there are specific moments (a story opening, a name mispronounced, the first fifteen seconds of a product experience) where my gut has to be the thing that decides. Me at my kitchen table with my four-year-old on my lap, listening. The machine cannot tell me where those moments are. I have to already know.

What Done Looks Like

The new job is not operator-of-agents. It is taste-bearer at scale: the person every piece of work has to pass through on its way out the door. The bandwidth I have is not for rewriting drafts. It is for knowing what good is and refusing everything below it.

If you are a tech worker reading this, the question is not whether agents will reshape your job. They already have. The question is whether you spend the saved capacity inside someone else's margin or on something you would not have built before. Creative-tech integration is the version I picked. There are others.

Pick one before the savings get reported instead of spent.

Try Endless Storytime · LinkedIn

*APRIL 24, 2026