Agents Made Every Interrupt Feel Free

A customer dropped a Slack: “I found this. Wouldn’t it be great if we could do that?”

I used to park those, write them down, and get back to the product that actually moved the number.

Now I paste it into an agent, because starting looks free. Twenty minutes later I have a branch, a passing test file, and a demo. I feel useful. I have not gone back to the bigger work. I have also not counted review, deploy, the defect this will throw in production, or the hour I will spend tomorrow explaining the shortcut to the next person who touches it.

I am watching a lot of good engineers do this. Laziness is the wrong diagnosis. They are available, and availability plus a cheap start is how the day gets stolen. Agents made every request feel like something you can finish before lunch. Generation hid the bill. The interrupt still collects it.

I have a hypothesis about who wins the next decade, and I want the argument on that. The park is not part of the bet, and you still owe it today.

Why Does Starting Look Free?

The old interrupt had friction. A request meant a day of typing, a design review, a fight for a sprint slot. That friction was ugly, and it also acted as a filter. Cheap ideas still cost enough that you asked whether they belonged.

An agent collapses the front of that cost. You can have a working copy of the flow before you have loaded the real system in your head. The demo is real, and the invoice stays off the screen.

Addy Osmani named the 80% problem in early 2026, building on Andrej Karpathy’s split of agent coding plus human touchups. Agents ship the visible eighty: the happy path, the screen, the test that is easy to write. They skip the other twenty: retries, audit logs, the place this feature lives in the rest of the product. Retrofitting that twenty later costs more than building it on purpose. Speed and mergeability go up while quality still erodes.

Martin Fowler drew a hard line around vibe coding: prompt, try it, prompt again, and do not look at the code. That can work for disposable software and it gets dangerous once other people have to live in it.

A 2026 paper called SlopCodeBench measured what happens when agents extend their own work across changing specs. Structural erosion rose in most trajectories. Verbosity rose with it. Agent code came out more bloated and more eroded than human histories of real Python repos.

The New Stack put a name on the layer under the slop: context debt. Armin Ronacher and Mario Zechner, who helped popularize agentic tools, told the Wall Street Journal the flood of bad code will catch up. Anand Asthana’s point is sharper. A regular agent can write a function and still cannot tell which of your seventeen payment APIs you are supposed to use.

TechTarget’s vibe-slop brief is blunt in a different way. Treat generated code as untrusted input. Keep a human owner who can explain, defend, test, and maintain it.

I agree with most of that. I also think they are describing the wreck after the crash.

The Interrupt Is the Selection Mechanism

The missing piece is how the work got chosen.

A developer is on the thing that should move the product. A customer, a PM, or their own itch arrives: I found this, wouldn’t it be great. They shove it into an agent because the start looks like net zero. They get lost in the side quest. They do not return to the bigger picture. Sometimes they interrupt themselves. Nobody pinged them. They saw a small ugly thing while loading the real system, pasted it because a cleanup looked like ten minutes, and spent the afternoon on a helper that the user never asked for. The Slack did not require a paste.

Aristotle called this akrasia (ἀκρασία): you know the better action and you do the other one. You know the product work is the work and you take the cheap request anyway. You picked the request and the model typed it.

The hidden bill is still the old bill: time to load the real system, review, deploy, the defect you will train the next person on, and the context you dumped to start the side quest. None of that is free. Generation hid it.

Then the product gets a new kind of damage. You keep the primary function and you bolt on a shortcut. The shortcut works in the demo and lives in neither the architecture nor the user’s path. You did not ask how the screen or the service fits the rest of the system. You locally optimized a request. Do that thirty times and you get slop: a product whose parts were started as separate yeses and no longer add up to one experience a user can finish.

People look at that pile and blame the machine. AI is writing shitty architecture. AI is writing shitty UX. I think the human is choosing the work. The model will generate whatever you start. If you start every Slack, you get a product that reads like a Slack archive.

Code got cheap is the cousin about judgment: typing is no longer the job. Big PRs is the cousin about review size. This one is about the moment before either of those. Did this request deserve to enter the shop at all?

If everything is urgent is the leadership version of alarm inflation. This is the individual version, where you became the alarm.

What I Think Wins

I keep coming back to a boring pair: discipline and patience.

enkrateia (ἐγκράτεια) is the struggle version: you feel the pull to paste and you hold. sophrosyne (σωφροσύνη) is the settled version, when the cheap request stops looking tasty. phronesis (φρόνησις) is the faculty that asks whether this is the right work for this day. kairos (καιρός) is the sense that this request might be right later and is wrong now.

The person who can sit in the bigger product while the Slack lights up will own the next decade. That is the forecast, and it can be wrong. The person who cannot will ship a lot of demos and a mediocre product that does not make the money. The demo is the appearance of progress, and the user’s week is the thing the work is for.

Waiting is not cover for a live outage, a promised ship date, or a broken path a user is on today. Those are the work, and the cheap start is the other pile.

Ignore is the wrong read, because the move is to stop treating a fast start as a cheap finish. Park the request on a list you will read at the end of the day, not in a new worktree. Name the bill out loud: load time, review, deploy, the defect, the abandoned product hour. Ask one systems question before you paste: where does this live in the architecture, and where does it live in the user’s week? If you cannot answer, you are not ready to generate. If the answer is “it lives nowhere, we would be inventing a side door,” you already have the decision.

You do not control how many “wouldn’t it be great” messages arrive, and you still control whether you treat the agent as a reason to start. That duty holds if you have no agent. Email and hallway requests had the same vice. The agent made the vice cheap enough to run all day.

A junior on a ticket mill still owes the park. A contractor on someone else’s backlog still owes the systems question. Missing seniority does not make the interrupt free. A manager who scores you on opened branches is circumstance, and the paste is still yours. arete (ἀρετή) is how you fulfill the function of the work, not how many branches you opened before lunch.

Tell Me Where This Breaks

I have seen people answer this three ways.

Some say the slop is the model, full stop, and the fix is a better model or a better scaffold. Work like TRIM tries to cut leftover edits from the agent’s own search. That work is real and it still does not decide which request should have been typed.

Some say the slop is the org: no product owner, no intake, no later pile. Also real. An engineer still assents to the paste.

Some say speed is the strategy and the mess is the tuition. Sometimes. A pile of incongruent shortcuts leaves a product nobody can hold in their head, including the next agent you hire to clean it.

I could be wrong in the weight of the forecast. Maybe generation quality is the main leak and the interrupt is a side story. Maybe the teams that win the decade will be the ones that take every request and get frighteningly good at cleanup. I do not think that is what I am watching. I think the filter moved from “can we type this” to “will we wait,” and most people have not noticed they now need a new muscle. Wrong on the decade still leaves the park owed this afternoon.

If you have a counter-example, I want it. A team that takes every Slack, ships, and still has a coherent product. A case where waiting was the slop. Send it.

Final Thoughts

Agents did not invent the cheap yes. They removed the last excuse for saying the yes would take too long.

Generation hid the bill. Review, deploy, defects, and the product you abandoned for twenty minutes still collect it. If we keep blaming the model for architecture and UX that a human chose to start, we will keep starting slop and calling it a tooling problem.

The engineer who can wait, stay on the bigger work, and think through the few requests they take, will eat the field. I think that is the hypothesis. I want the argument.

Ready to train the muscle that stays on the work that matters? Join us at MasteryLab.co.